EUROPEAN COMMISSION

Brussels, 28.9.2022

SWD(2022) 319 final

COMMISSION STAFF WORKING DOCUMENT

IMPACT ASSESSMENT REPORT

Accompanying the document

Proposal for a Directive of the European Parliament and of the Council

on adapting non-contractual civil liability rules to artificial intelligence

{COM(2022) 496 final} - {SEC(2022) 344 final} - {SWD(2022) 318 final} - {SWD(2022) 320 final}

Table of contents

1.Introduction: Political and legal context

2.Problem definition7

2.1.Introduction and overview7

2.2.Problem driver: the specific characteristics of AI make it difficult to meet the burden of proof

2.3.Problem 1: Legal uncertainty leading to internal market obstacles0

2.4.Problem 2: Legal fragmentation leading to internal market obstacles1

2.5.Problem 3: Lack of compensation for victims leading to a lack of societal trust and hampering uptake of AI-enabled products and services2

2.6.The consequences: internal market obstacles and affected stakeholders

2.7.Likely evolution of the problems without Union intervention

3.Why should the EU act?19

3.1.Legal basis19

3.2.Subsidiarity: Necessity and added value of EU action

4.Objectives: What is to be achieved?

4.1.General objectives

4.2.Specific objectives

5.What are the available policy options?

5.1.What is the baseline from which options are assessed?

5.2.Description of the policy options29

5.3.Measures discarded at an early stage

6.What are the impacts of the policy options?

6.1.PO1: Easing the burden of proof for AI-related claims2

6.2. PO2: PO1 + strict liability for AI use-cases with a specific risk-profile, possibly coupled with mandatory insurance ……………………………………………………………………….. 51

6.3.PO3: PO1 + targeted review regarding the strict liability and mandatory insurance elements of PO2 (staged approach)57

7.How do the options compare?59

8.Preferred option60

8.1.Rationale and main impacts of the preferred PO60

8.2.The interplay with the PLD review: consistency and complementarity of the package of measures for the compensation of damage caused by AI61

8.3. The interplay with the PLD review: consistency and complementarity of the package of measures for the compensation of damage caused by AI ………………………………….. 61

   8.4.    One-In-One-Out …………………………………………………………………………… 62

9.How will actual impacts be monitored and evaluated?62

Annex 1: Procedural information

Annex 2: Stakeholder consultation

Annex 3: Who is affected and how? 93

ΑΝΝΕΧ 4: Analytical methods and key findings of supporting studies / Explanations on how the European Parliament’s legislative own-initiative resolution on a civil liability regime for AI was taken into account 101

ΑΝΝΕΧ 5: DETAILED EXPLANATIONS ON HOW THE SPECIFIC CHARACTERISTICS OF AI ARE CHALLENGING EXISTING LIABILITY RULES ……………………………………….…..... 120

Annex 6: Detailed description of the legal context and interplay between existing / proposed legal rules and the AI liability initiative

Annex 7: Interplay with the AI Act

Annex 8: AI-specific fundamental rights concerns – Overall Commission policy approach and the role of the AI liability proposal

Annex 9: The types of compensable (in particular immaterial) harm and the admissibility of contractual exclusions/limitations of liability – Member States’ legal approaches/traditions and reasons for not harmonising these aspects specifically for AI

Annex 10: Detailed explanations and results regarding the assessment and comparison of policy options (multi-criteria analysis)

Annex 11: The economic impact of adapting civil liability rules to the specific challenges of AI – The example of the European market for vacuum cleaners

Annex 12: Monitoring and evaluation

Annex 13: Illustration of AI-specific difficulties in claiming compensation based on case scenarios

1.Introduction: Political and legal context

1.1. Introduction and political context

AI is a set of enabling technologies with the potential to transform our economy and society to such a profound extent that it is often compared to electricity. As a key driver for future economic development, AI has strategic importance 1 . While AI can bring about economic and societal progress, it can also put at risk European values.  In particular, when AI systems cause harm, the victims’ legal interests of life, health and property are at stake. In addition, the right to an effective access to justice can be in doubt because the specific characteristics of certain types of AI can make it prohibitively difficult to prove their claim under existing liability rules. 

For these reasons, Commission President Ursula von der Leyen announced in her political Guidelines 2 a coordinated European approach on AI. In its White Paper on AI 3 , the Commission undertook to promote the uptake of AI and to address the risks associated with certain of its uses by building an ecosystem of excellence and trust. One of the White Paper political objectives for liability and this impact assessment is that persons having suffered harm caused by AI systems enjoy the same level of protection as persons having suffered harm caused by other technologies. 4 In the Report on AI Liability 5 accompanying the White Paper, the Commission identified the specific challenges posed by AI to existing liability rules. The Commission Work Programme 2020 envisaged a follow-up to the White Paper in the form of new legislative initiatives, including on liability. 6

In its conclusions on shaping Europe’s digital future of 9 June 2020, the Council welcomed the consultation on the policy proposals in the White Paper on AI and called upon the Commission to put forward concrete proposals. On 20 October 2020, the European Parliament (EP) adopted a legislative own-initiative resolution requesting the Commission to adopt a proposal for a civil liability regime for artificial intelligence based on Article 114 TFEU. 7

The Coordinated Plan on AI in the Annex to the Communication ‘Fostering a European approach to Artificial Intelligence’ of 21 April 2021 8 specifies that the Commission will propose in 2022 measures adapting the liability framework to the challenges of AI.

AI systems could also contribute to the accomplishment of a several targets across the Sustainable Development Goals included in the 2030 UN Agenda for Sustainable Development. 9 The use of AI could support environmentally beneficial outcomes for instance by improving prediction as well as optimising operations and resource allocation.

Liability-related issues linked to AI are the subject of debate in various EU trading partners. As a consensual regulatory solution has not yet emerged, the EU has the opportunity to influence the global discussion with a practical and constructive approach which has a chance to become a global model.

In terms of strategic foresight, the AI liability initiative takes into account the megatrend of Accelerating technological change and hyper-connectivity. This initiative is a crucial part of the EU’s efforts to shape the AI-driven transformation in a human-centric, yet innovation-friendly way. The potential of AI underlines the importance of creating the right conditions for its rollout through legal certainty and consistent rules across the internal market. The ubiquitous and fast evolving nature of AI also means that the risks linked to the deployment of AI systems can potentially affect a very broad range of stakeholders and legal interests. By adapting the traditional liability rules to the challenges of AI, the initiative anticipates the novel difficulties of proof faced by victims of harm, and aims to ensure that societal trust in AI is not diminished due to a lack of compensation in the future. Due to this forward-looking perspective of the AI liability initiative, a strategic foresight approach underlies the problem definition, policy objectives and policy options.

1.2.Scope of the impact assessment: AI systems and their characteristics

In order to ensure consistency, it is appropriate to use the same general concept of AI as in the AI Act 10  also as the basis for the present initiative which is complementary to the AI Act.

AI-specific issues regarding liability are linked to certain characteristics of AI systems (opacity/lack of transparency and explainability, autonomous behaviour, complexity, continuous adaptation and lack of predictability) 11  which make the application of existing liability rules uncertain or more difficult (see the following section on legal context). AI systems not having such characteristics can be dealt with under the existing liability rules, similarly to other types of software. For example, rule-based algorithms are designed to automatically execute rules encoded by their programmers, following the classical ‘if → then logic’ as other types of software. Such algorithms do not put victims of harm in a fundamentally more difficult situation than other types of software. For example, robotic process automation in a factory setting may rely on such rule-based systems. In contrast, the highly autonomous operation of e.g. mobile robots in more unpredictable environments (such as delivery robots) – often involving interaction with humans – will typically require more complex AI systems, in particular developed through machine learning. Such AI systems are expected to challenge existing liability rules and are therefore concerned by the present initiative. 

Their operation can potentially affect legally protected interests (for example AI used for recruitment, in autonomous vehicles, delivery robots, or drones). Certain AI-enabled products and services with the potential to cause harm are already in an advanced development stage or even starting to be deployed, such as e.g. certain mobile robots or AI-systems used in the context of recruitment. However, AI systems with the specific characteristics challenging liability rules are mostly not yet on the market, among others because they are not yet approved. Therefore, there are no available judicial cases yet where the problems to be addressed by this initiative have materialised. The objective of this proposal is to create the legal certainty conditions and to promote the societal trust needed for such AI systems to be brought to market in the EU.

1.3.Legal context 12

Liability rules determine how damage, caused by human activities and goods for which humans are considered liable by law, can be compensated.

This initiative concerns ‘extra-contractual’ civil liability rules, i.e. rules providing a compensation claim irrespective of a contractual link between the victim and the liable person. Those rules will be referred to as ‘liability’ in the following. Criminal liability is not concerned.

A victim currently relies on national liability rules and in certain cases on the Product Liability Directive 13 (PLD) to claim compensation for damage arising from AI-enabled products and services. 

Depending on each case, a victim has several avenues available for claiming compensation. Various persons could be held liable in parallel under different conditions and for different reasons. Neither of these rules contains specific provisions for AI.

Firstly, the victim could have a fault-based liability claim based on another person’s wrongful behaviour, i.e. fault. 14 A fault-based claim usually requires proving the existence of:

-a damage,

-a fault of the liable person, and

-causality between that fault and the damage.

As a matter of principle, it is the victim that needs to prove these three elements. 15  

Such claims apply irrespective of the technology/activity involved in causing harm and can be addressed against any type of wrongdoer (e.g. businesses, including service providers, or private persons) by any type of victim (e.g. business, private individual, public entity). They can compensate any type of harm protected by national law, e.g. physical injury, property damage, damage to or loss of data, economic loss like lost profit, non-economic loss like pain/ suffering or harm caused by discrimination. 16  

Secondly, in addition to fault-based claims, victims could rely on liability independent of fault - so-called strict liability, with different scopes and conditions set by national laws. 17  

Strict liability rules assign liability for the relevant risk to a person, irrespective of fault. This is usually justified by the fact that this person draws a benefit from exposing the public to that risk 18 . Normally, for that person to be found liable, the victim has only to prove that the risk stemming from the sphere of the liable person materialised.

Thirdly, victims may have, under the PLD, a claim against the producer of a defective product for a defect present when the product was put on the market. The victim must prove that the product was defective and the causal link between that defect and the damage.

Under the PLD, the victim can claim compensation for personal injury or damage to consumer property. The PLD does not cover damage caused to property intended for professional use, damage caused during the provision of a service, damage to victims other than natural persons or claims based on the wrongful use of a product. In all these cases, victims can only get compensation according to national liability rules.

Types of civil liability

1.4.Related policy initiatives

The Commission is aiming to support the roll-out of AI in Europe by developing an ecosystem of trust. In accordance with the White Paper, the Commission proposes a forward-looking comprehensive package of measures to address problems caused by the introduction and use of AI, comprising of three complementary work streams:

-a horizontal framework addressing fundamental rights and safety risks specific to AI systems (AI Act); 19  

-a revision of sectoral and horizontal product safety rules;

-EU rules to address liability issues related to AI systems.

Interaction with already adopted proposals: The Commission has already proposed rules in the AI Act reducing risks for safety and fundamental rights. While they will reduce risks, they are not intended to prohibit the placing on the market of AI-systems posing a residual risk to safety and fundamental rights. 20  Harm will therefore still occur. When this happens, the liability rules subject to this impact assessment determine which party is liable for harm and under which conditions a victim can be compensated. 

The AI Act also provides for definitions and specific documentation/logging requirements for high-risk AI systems. The information from these documents and logs could be useful in liability claims but the AI Act does not foresee a right of victims to access that information. 21  

The Commission has also put forward updates of general 22  and sectoral product safety rules applicable to AI-enabled machinery products 23 and radio equipment. 24  These initiatives do not contain provisions on liability or access to information for the purposes of damages claims. Their requirements (e.g. general safety requirement, risk assessment or regarding human oversight, testing, data quality) can be relevant for determining what constitutes fault of producers or what is a defective product in relation to AI – points which the AI liability proposal will not regulate. 

Safety and fundamental rights requirements are meant to prevent, monitor and address risks and thus address societal concerns 25 . While they employ tools that address a potential wrongdoer, such as authorisations, controls, monitoring, administrative sanctions, they do not compensate the victim for the harm suffered.

Therefore, safety and liability are two sides of the same coin: they apply at different moments and reinforce each other. Safety rules intend to reduce the risk of harm. Liability rules come in to ensure compensation of the victim when harm has nevertheless occurred. Effective liability rules also provide an economic incentive to comply with safety rules and contribute therefore to preventing the occurrence of harm. 26

As AI is only gradually being rolled-out, there are not yet many documented cases of harm as the relevant products/services are not yet on the market. However, as part of the Commission’s forward-looking and enabling approach for the uptake of this technology, action is needed to adapt the applicable liability rules in advance in order to increase trust and create the conditions for the roll-out of these products and services.

Interaction with the pending revision of the PLD: The Commission takes a holistic approach in its AI policy to liability by considering adaptations to the producer’s liability for defective products under the PLD as well as a targeted harmonisation of specific aspects of the other liability rules.

These two policy initiatives are closely linked and form a package, as the claims within their scope deal with different types of liability. They complement one another to form an overall effective civil liability system. For AI, they achieve the AI White Paper objective that persons having suffered harm caused by AI systems enjoy the same level of protection as persons having suffered harm caused by other technologies.

Together they will deliver the ex-post trust element in AI (and other digital technologies), by ensuring that victims have an effective chance of compensation if damage occurs despite the preventive requirements of AI Act and the other safety rules.

2.Problem definition

2.1. Introduction and overview

Liability ranks amongst the top three barriers to the use of AI by European companies in a representative survey of 2020. It is the most relevant external obstacle (43%) for companies that are planning to, but have not yet adopted AI. 27  In the White Paper, the Commission counted liability-related issues amongst the main risks related to the use of AI, because victims may not have effective access to the evidence necessary to build a case in court and economic operators face uncertainty as regards the allocation of responsibilities.

The accompanying Report identified those AI-specific characteristics that challenge liability rules, (see section 2.2.). Economic 28 , behavioural 29 and legal 30  studies commissioned for this impact assessment, as well as comprehensive stakeholder consultations 31 , have shown three AI-specific problems: legal uncertainty (see 2.3.), increased legal fragmentation (see 2.4.) and a lack of compensation for victims contributing to the lack of societal trust in AI use (see 2.5.), which in their combination cause internal market obstacles (see 2.6.).

Delineation from the problems covered by the parallel impact assessment on the PLD review: Under the PLD, the burden of proof is less onerous on victims than under national fault-based liability rules, because the victim does not have to prove fault. Nevertheless, proving that a product is defective (i.e. that it does not provide the level of safety a person is entitled to expect) is challenging for complex products, of which AI-enabled products are an example, even if the victim does not have to prove how the product became unsafe.

As the PLD regulates producer liability for various different types of defective complex products, of which AI-enabled products are only a part, this problem is covered in the parallel impact assessment on the PLD review. It will be addressed by amendments of the PLD regulating in a horizontal manner all those products.

2.2. Problem driver: the specific characteristics of AI make it difficult to meet the burden of proof

The three problems (uncertainty, fragmentation and lack of compensation) are driven by the fact that, under existing liability rules, victims normally have to prove the necessary conditions for a successful claim. Under current rules, victims generally have to identify the possibly liable person(s), the action/omission of that person which could be characterised as fault, and the causal link between that action/omission and the damage. There are neither targeted provisions for AI to ease the victim’s burden of proof under fault-based claims, nor AI-specific strict liability regimes that would exempt the victim from making this proof.

However, certain AI-systems have characteristics that make it prohibitively difficult or even impossible for victims to provide that proof. 32  The specific characteristics that challenge current liability rules are: increasingly autonomous behaviour, opacity/lack of transparency and explainability, complexity, continuous adaptation and lack of predictability. 33 These characteristics make it very difficult to attribute directly the damaging output of the AI to an action/omission of a liable person, i.e. to identify:

-the person who could possibly have done something wrong,

-the action or omission of that person which might not comply with the relevant standard of care, and

-the causal link between that action or omission and the victim’s damage.

This is linked to two situations that can occur due to the use of AI systems:

-between the human actor(s) related to the AI system and the damage caused by this AI system to a victim, the AI-system takes an autonomous action which is not pre-determined by a human and/or

-the way the AI action is performed cannot be understood (mainly due to the opacity and complexity of certain forms of AI).

However, in situations where the liable person acted on advice or recommendations given by an AI system, this AI-specific problem driver does not apply. In such cases, the human acting on the advice or recommendation will be responsible as there is a human action/omission, which can be a) identified, b) characterised as not complying with the relevant standard of care, and which would c) be the cause for a specific damage. This is for example the case with AI systems providing medical analysis or even suggestions for diagnosis and treatment, which are feeding into a decision on diagnosis and treatment, but that decision is ultimately taken by a human physician. 36  

As AI-enabled products and services with the specific characteristics challenging liability rules are for the most part not yet on the market, it is not possible to quantify the scale of the problem or of its relevance for different market segments. The economic study did however provide some indications. 37 The market share affected by legal uncertainty and legal fragmentation regarding AI liability will likely differ substantially per sector and is potentially larger where AI technologies directly interact with humans and where – in the case of a failure or damage – large-scale negative effects can be expected. Potential bodily and psychological harm, property damage, economic loss and damage caused by discrimination present recurring risks linked to the use of AI across various sectors. 38 The economic study found 39 that the share of AI-enabled products and services potentially affected by legal uncertainty and legal fragmentation of civil liability rules ranges between ca. 20% in sectors such as real estate or oil and gas and around 40% in sectors such as human health and transport.

2.3. Problem 1: Legal uncertainty 

As there are currently no rules addressing the AI-specific difficulties of proof, the general burden of proof conditions apply. This has several consequences, creating legal uncertainty.

The victim and the possible liable person wanting to check liability risks, as well as the judges having to decide a liability claim, will be confronted with cases where they would need to interpret general rules which were not designed with AI in mind. They would have to apply them to AI-technologies that are qualitatively different from previous technologies, because of their specific characteristics (see 2.2.). It is quite uncertain how existing liability rules would be interpreted in light of the AI-specific difficulties of meeting the burden of proving fault and causation. 40  

Such difficulties would lead in all likelihood to the result that the victim would not be able to meet this burden of proof. Victims may therefore decide to not even claim compensation. As an economic consequence, they would have to bear the burden of the damage.

If the case goes to court nevertheless, national courts may decide - on an ad-hoc basis - to stick to the traditional allocation of the burden of proof on the victim. This would lead in all likelihood to the result that the victim would lose the case, not obtain compensation and bear the economic burden of the damage. If such a result is not considered as equitable, national courts may decide - again on an ad-hoc basis - to adapt to the specificities of AI and alleviate the burden of proof of the victim, in order to achieve a fair outcome in each single case.

These uncertainties surrounding AI liability are especially relevant in a cross-border context, because MS’ existing burden of proof rules are already very diverse. 41  The economic impact of this legal uncertainty is demonstrated by legal experts. When asked to estimate the costs of claiming compensation based on national liability rules for damage caused by AI, they submitted wide ranges to account for the prevailing uncertainty. 42

In the PLD context, unclear rules, including for AI-enabled products, have equally been identified as a problem driver. This mainly concerns the problem that the PLD is ill-adapted to products in the digital age in general. It imposes neither liability on software providers, including AI-system providers, nor liability for defects emerging after a product is placed on the market, such as through software updates, or ancillary digital services. Legal uncertainty applies horizontally to defective intangibles like software, digital content and ancillary services. For producers’ liability for defective products, including with respect to AI-enabled products, this problem and the measures needed to address it are discussed in the parallel impact assessment on the PLD revision.

2.4. Problem 2: Legal fragmentation

In a cross-border context, the law applicable to a liability claim is by default the law of the country in which the damage occurs. 43  This means that different liability regimes and burden of proof rules could be applied to the same kind of AI-enabled product/service deployed in several MS which causes the same kind of damage. 44  Consequently, a company disseminating/operating the same AI-enabled product/service in various MS could face different liability exposures if that product or service causes, for the same reason, the same type of damage in several MS. 45  Aside from the general burden of proof rules, there are also different rules about information disclosures in liability claims. This means that victims face greater information asymmetries in some MS. 46

In addition, national AI strategies show that several MS are considering, or even concretely planning, legislative action on civil liability for AI. 47 The fact that in the public consultation on the AI White Paper, more than 63 % of respondents were in favour of adapting national liability rules for all or for specific AI applications suggests that MS are likely to face pressure to address the issue of liability for damage caused by AI. 48

While the fragmentation of MS’ existing liability and burden of proof rules is as such not limited to cases involving AI, the likely adoption of AI-specific liability rules by at least some MS exacerbates the problem of fragmentation with respect to AI-enabled products and services in particular. Outside the scope of the PLD, tort law is currently not harmonised and MS’ rules diverge substantially. This may create internal market obstacles also for traditional products and services. Nevertheless, it is crucial to address this issue specifically for AI, for the following reasons:

-The Commission aims to create the right conditions specifically for the roll-out of AI-equipped products and services.  According to the Commission’s ‘staged approach’, this involves two complementary and synergetic initiatives: in a first step, the AI Act proposal to ensure the effectiveness of safety and fundamental rights protection, and in a second step, the AI liability proposal to ensure effective protection when AI nevertheless causes harm.

-Stakeholders active in the field of AI will be especially affected by the market barriers associated with legal fragmentation (see 2.6.).

-As a key enabling technology, the importance of which can be compared with the introduction of electricity, AI bears a significant potential for economic growth and societal usefulness. It is therefore essential to address the obstacles entailed by legal fragmentation in order to reap the full potential of AI within the internal market and for the digital transition.

-Given the entrenched and longstanding differences between national legal traditions, any attempt to address the problem of legal fragmentation by way of a broad stroke harmonisation of tort law would not be promising. Focussing on AI specific problems and addressing them with targeted adaptations to specific aspects of national laws provides an opportunity to approach the problem of fragmented civil liability rules in an effective and future-oriented manner.  

The existence of diverging approaches to easing the burden of proof in cases involving complex products also forms part of the problems affecting the PLD, as it leads to differing levels of consumer protection (section 2.2.2. in the PLD impact assessment). Since the PLD problems concern not only AI-enabled products but also other complex products, they are dealt with in the impact assessment on the PLD revision.

2.5. Problem 3: Lack of compensation for victims leading to a lack of societal trust and hampering uptake of AI-enabled products and services

As already explained (see 2.2.), some characteristics of certain AI systems challenge the application of liability rules and may make it prohibitively difficult or even impossible for victims to get compensation. Victims are likely to incur significantly higher up-front costs and face significantly longer legal proceedings, when claiming compensation for damage caused by AI-enabled products or services, compared to cases not involving AI applications. 49  Substantial burdens on victims are anticipated, due to the specific challenges of AI, for almost all national legal systems covered by the analysis commissioned for this impact assessment, often going beyond a 100 % increase of fees paid for legal and technical expertise. 50  This could deter victims from claiming compensation. Ultimately victims may be left under- or even uncompensated. AI-specific barriers are therefore likely to impinge on the right to an effective judicial remedy. 51 As a consequence, the compensation function of liability law would fail. This would undermine the fundamental concept of justice as fairness (‘distributive justice’) that underpins our legal order and citizens’ trust in the justice system.

If victims are left without compensation when AI is involved, they would find themselves at a disadvantage compared to a situation where traditional technology would have caused the damage. Such an AI-induced lack of compensation would be particularly problematic if future autonomous technologies would create a significant risk of harm to life, health and significant property values, due to the important legal interests at stake. The need to ensure effective compensation is all the more pressing where unwitting third parties are exposed to such risks, as such victims cannot choose to accept or avoid the risk of damage.

A behavioural study commissioned for this impact assessment has shown that the expected lower likelihood of receiving compensation if AI applications cause damage is amongst the reasons for low levels of consumer trust in such applications. 52  Conversely, the availability of effective remedies and compensation is one of the factors increasing the level of societal acceptance of, and consumer trust in, AI-enabled products and services. 53 Another recent survey confirmed that consumers have clear concerns regarding the allocation of responsibility and liability if something goes wrong. It concluded that these concerns must be properly addressed to ensure strong protection for consumers and for them to trust this technology. 54 A statistically representative Eurobarometer of 2019 points in the same direction: EU citizens’ most widespread concern about the use of AI was that it could cause situations where it is unclear who is responsible. 55  

The aforementioned behavioural research demonstrated that a lower level of trust in AI applications and the perception that persons suffering damage caused by AI-enabled products and services are less likely to obtain compensation correlate with a lower willingness to take up such products and services. 56 Businesses also consider that the lack of citizens’ trust is one of the most relevant obstacles to the take-up of AI-technologies in Europe. 57 Other factors, such as the perceived risk of accidents, were also shown to have a significant effect on the level of societal trust and consumer uptake. These findings are consistent with the complementary role of the AI Act / safety legislation and liability rules in creating an ecosystem of trust for AI.

2.6. The consequences: internal market obstacles and affected stakeholders 

(a)     Directly affected stakeholders

Potentially liable companies: The representative 2020 IPSOS survey of European companies showed that liability for potential damages is the most recurrent barrier to the adoption of AI technologies. 58  The economic study found that a majority of businesses active in the market for AI-enabled products and services perceive their liability exposure outside the scope of the PLD as uncertain to varying extents. 59  

Where companies are considering a cross-border rollout of AI-enabled products or services, the costs and burden linked to legal uncertainty are amplified by fragmentation between MS’ liability rules. This means that such companies have to carry out complex assessments and quantifications of liability risks for each relevant MS. A significant share of businesses active in the AI market does or will incur costs due to legal fragmentation and uncertainty regarding the application of national liability rules to damage caused by AI: 60 legal information/representation and compliance costs, internal risk management costs, opportunity costs in the form of foregone revenue due to companies’ hesitation to explore new markets. 61  

Moreover, liability rules may sometimes have a bearing on which technological solution is favoured by developers of novel products and services 62  and may influence how products and services are designed. 63 Going cross-border may therefore involve costly technological adaptations that thwart economies of scale. Companies deterred from investing in AI-technologies miss out on benefits in terms of increased profits or efficiency gains. 64  

The economic study estimated that the total values of the EU27 AI market affected by legal uncertainty and legal fragmentation ranges from EUR 1.739 billion to EUR 4.973 billion in 2021. 65  These shares of the total AI market in the EU27 were obtained by multiplying the value of the AI market per sector with the respective shares affected by legal uncertainty and fragmentation. To ensure consistency, the estimated AI market sizes assumed for the impact assessment accompanying the proposed AI Act were taken as a starting point, representing overall investments in AI technologies. 66  For the sake of robustness and in order to account for uncertainty, these estimates allow for a wide range of market definitions and possible future developments. 67 The shares affected by legal uncertainty and fragmentation were determined based on a sector-by-sector analysis, taking into account the ability of the AI applications deployed in the respective sector (now or in the foreseeable future) to cause harm and thus give rise to liability claims. 68    

The EP concluded in 2020 that fragmentation and uncertainty, relating to the liability provisions applicable across the EU, that may emerge in the absence of a common EU approach can provide negative incentives for innovation and diffusion of AI systems while also contributing to excessive costs for consumers. 69  These internal market obstacles will likely contribute to slowing down the innovation that drives the overall digital transition. Legal uncertainty and fragmentation regarding civil liability presents significant barriers across all sectors 70  and both to products and services. 71

SMEs in particular: Start-ups and other SMEs are likely to be significantly more affected by legal uncertainty and fragmentation than large companies: they cannot rely on comparable in-house legal expertise and an established cross-border network. As smaller companies often lack sufficient external funding, the additional costs incurred due to legal uncertainty may be prohibitive for start-ups and other SMEs, or at least reduce their competitiveness in relation to large companies. 72  These effects are particularly weighty because the EU market is driven by SMEs developing, deploying or using AI technologies. It is estimated that SMEs account for more than 95 % of companies active in this market, with micro-enterprises representing over 80 % of all firms involved in AI research and software development. 73  Yet it is estimated that only 19 % of micro-enterprises active in the internal AI market are involved in cross-border activities (as opposed to 68 % of large corporations). 74  This points to a significant unused internal market potential. 75 As most start-ups and other SMEs are not using the potential of the single market to scale up, the fragmentation of the AI landscape, notably in terms of regulations, impedes the creation of a regional ecosystem with bigger scale. 76

Insurance companies: Legal uncertainty regarding the liability exposure of insurance holders can make it more difficult to calculate premiums and offer coverage, in particular for cross-border activities and for situations covered by fault-based liability. 77 Such difficulties affect the insurance companies themselves. But they are also of concern to the companies facing unpredictable – and potentially very high – liability exposure when unable to procure desired insurance coverage. Finally, they also affect the victims bearing the liable party’s insolvency risk. Moreover, liability insurance premiums depend, to some extent, on insurers’ prospect of successful redress claims allowing them to claim back the amounts they paid out. 78 Outside the scope of the PLD, such redress claims rely to a large extent on fault-based liability, which means that insurance companies will face significant uncertainty in estimating their own chances of success on the redress level for damage caused by AI-enabled products and services.

Victims of damage caused by AI-technologies may be (i) private persons, businesses or public entities owning AI-enabled products or using AI-enabled products and services and (ii) natural persons, businesses or public entities without any link to those products and services (third-party damage). See 2.5. for detailed explanations on the difficulties faced by victims.

(b)     Indirectly affected stakeholders

Consumers: As explained under 2.5., a lack of compensation of damage caused by AI is likely to contribute to low levels of consumer trust and demand of AI-enabled products and services. 79 If restrained consumer demand, coupled with legal uncertainty and fragmentation faced by businesses, delay the roll-out of novel AI-technologies in the internal market, consumers would, to some extent, be deprived of the benefits promised by AI. They could miss out on benefits like faster and more personalised services, innovative and performant products as well as advances in the fields of health, safety, security, mobility, sustainability, media, etc.

Companies active in AI: The trust deficit and the consequential lower uptake of AI-enabled products and services is likely to affect all companies with a stake in the rollout of AI in the form of lost market opportunities. This includes also those entities in the AI value chain, such as researchers, developers and providers of related technologies or equipment (e.g. cloud service providers), who are likely to face reduced demand.

General economic consequences: If victims are left to bear the burden of damage caused by AI, the overall cost allocation of introducing this technology is inefficient, as this cost is borne by those that neither can control the risk nor benefit from it. Burdening victims disproportionately and allowing for an externalisation of costs linked to the roll-out of AI would create inefficient incentives. It would also undermine the innovation potential linked to the development of safe AI-enabled products and services. 80  

In an even broader sense, the existence of obstacles to the functioning of the internal market in AI-enabled products and services affects the European economy and society as a whole. The cross-sector relevance of AI was confirmed by a representative survey of 2020, showing that at least 35 % of EU companies from each of the sectors covered have adopted AI technologies. 81  The identified liability-related obstacles are likely to have a negative impact on the overall growth and innovation capacity of the European market. The fact that EU companies cannot take full advantage of the internal market prevents them from becoming more competitive on a global scale, and may thus further weaken the EU’s position vis-à-vis its competitors in the global AI race (primarily the US and China), as the EU already lags behind them 82 . 

By putting forward a balanced approach regarding liability for damage caused by AI, the EU has the opportunity to set a global benchmark and ultimately generate a competitive advantage for ‘AI made in Europe’.

While the longer term impact of Covid-19 on the situation of the internal market for AI-enabled products and services is still somewhat uncertain, the pandemic has expedited the trend towards increased digitalisation, and economic growth has picked up strongly in 2021 and 2022. 83  The war in Ukraine is likely to have a significant negative effect on the EU economy, and forecasts will have to be reviewed taking into account his factor. However, at the time of submitting this impact assessment no updated estimates were available yet.

General societal consequences: A liability framework which is significantly more difficult to apply when AI is involved will not sufficiently motivate risk takers (producers, service providers, operators, users, etc.) to make efforts to avoid harm to others. In particular, inefficient liability rules will not sufficiently motivate compliance with safety rules or with the ex-ante requirements introduced by the AI Act.

Compensation gaps could call into question citizens’ trust not only in AI, but also in the ability of the legal and judicial system to ensure fair and equitable results.

2.7. Likely evolution of the problems without Union intervention

Legal uncertainty regarding the application of national liability rules might to a certain extent diminish in the long term as MS adopt AI-specific liability rules or national courts clarify the interpretation of existing rules. However, the adoption of AI-specific liability rules at the national level would in all likelihood exacerbate the problem of fragmentation since such rules would be different and not benefit from a harmonised interpretation by the European Court of Justice.

Judicial solutions adopted by national courts are by nature limited to the particular case at issue and in most cases do not bind other courts, which in turn can adopt different solutions. Therefore, legal uncertainty for business and victims’ difficulties in claiming compensation will persist, to a large extent, while fragmentation in the internal market will increase.

The AI Act and various safety related rules 84  aim to improve the availability of relevant information on, and the transparency and explainability  85 of high-risk AI systems, the safety of those systems and the effective protection of fundamental rights. The requirements of the AI Act and future detailed standards providing details on these requirements can help to determine the duties of care for fault-based claims. The information to be documented and the logged records of the system pursuant to the AI Act can be useful, if the victim has access to it, for establishing liability. However, these proposals do not contain provisions designed to facilitate the victim’s claim, e.g. on the burden of proof or access to information. They will therefore not remove the AI-specific problem drivers in the context of liability. For this reason, the impact assessment for the AI Act underlined the importance of combining that initiative with specific liability-related measures.

It is to be expected that the problem drivers of the difficulty of claiming compensation linked to e.g. the autonomy, opacity and complexity of certain AI systems will become more relevant in the future, as increasingly independent AI-systems will carry out ever more advanced tasks (see 5.1.).

The negative effect of a perceived lack of compensation on societal trust is likely to further increase, if the risks posed by AI-enabled technologies materialise in practice and the public becomes increasingly aware of difficulties in claiming compensation. Accidents involving AI-enabled products often receive a very large amount of public attention. 86  Behavioural research has shown that the level of consumers’ anxiety towards the use of AI applications and their likelihood to take up AI-enabled products/services is correlated with their exposure to media coverage of accidents involving such products/services. 87  This effect could already be observed e.g. by the change in public perception of self-driving vehicles following high-profile crashes. 88  It is very likely to become more relevant with the increased rollout of AI applications capable of causing harm. 

As the market for AI-enabled products and services is expected to grow substantially, and AI can be implemented in more and more situations and sectors, the economic and societal relevance of the identified problems is also likely to increase in absolute terms. 89  The economic study concluded that the sum of the AI market share affected by liability issues will increase at a compound annual growth rate of 44% – 56% until 2025. 90  

3.Why should the EU act?

3.1. Legal basis

The initiative constitutes a core part of the EU AI strategy. The problems identified in section 2 (legal uncertainty, legal fragmentation, lack of compensation diminishing consumer trust) hinder the development of the internal market and thus amount to significant obstacles to cross-border trade in AI-enabled products and services. 91  

On the one hand, these obstacles stem from the fact that businesses that want to produce, disseminate and operate AI-enabled products and services across borders are uncertain whether and how existing liability regimes would apply to damage caused by AI. This uncertainty concerns not only their own MS but particularly other MS where they will export to or operate their products and services. This assessment was recently confirmed by the comparative law study. 92  Certainty would be necessary for these businesses to know the relevant liability risks and to be able to insure themselves against them.

On the other hand, there is concrete evidence showing that a number of MS would take unilateral legislative measures to address the specific challenges posed by AI with respect to liability. For example, AI strategies adopted in Italy 93 , Malta 94 , Czech Republic 95 , Poland 96 and Portugal 97 mention initiatives to clarify liability. Given the already existing large extent of differences between MS’ existing civil liability rules, it is very likely that any national AI-specific measures on liability would follow existing different national approaches and therefore increase such differences.  

In a cross-border context the law applicable to a non-contractual obligation arising out of a tort/delict is by default the law of the country in which the damage occurs. 98  Thus, adaptations of liability rules taken on a purely national basis would increase the barriers to the roll-out of AI-enabled products and services across the internal market and contribute further to fragmentation. 

In light of the existing situation and the content and aim of the planned initiative, the appropriate legal basis is Article 114 TFEU. This initiative addresses the abovementioned obstacles by harmonising targeted aspects of MS’ existing civil liability rules applicable to AI-systems, in order to improve the conditions for the functioning of the internal market in AI-enabled products and services. The choice of that legal basis is also supported by the EP which has called twice upon the Commission to use this legal basis for a legislative proposal. 99

3.2. Subsidiarity: Necessity and added value of EU action

(a) Necessity of action at EU level

The identified obstacles have significant impact on cross-border trade of AI-enabled products and services. The negative effects of legal uncertainty – additional legal information/representation and risk management costs as well as foregone revenue – were found to be worse for companies trading cross-border. 100 Likewise, fragmentation between MS regarding the conditions under which business would face compensation claims for damage caused by AI would increase transaction costs especially regarding cross-border trade, entailing significant internal market barriers. 101  

The significance of these obstacles is reinforced by the fact that legal uncertainty and fragmentation disproportionately affect start-ups and other SMEs, which account for the large majority of companies and the major share of investments in the relevant markets. 102 Smaller businesses lack the necessary financial resources as well as the specialised legal and technical expertise to cope with the identified challenges. Thus, they are more likely to be deterred from rolling out their products or services across borders 103 , incurring lost revenue. This situation is making the EU market overall less strong and innovative and European companies less competitive.

MS could to some extent address the problem of legal uncertainty and improve the effectiveness of liability claims by adapting their respective rules. However, they would necessarily do so by building on the approaches of their own, already very different liability regimes. This would lead to further fragmentation, with the negative effects outlined under 2.4. and 3.1. The relevant problems can therefore be addressed effectively only at EU level.

(b) Added value of EU action

The conditions for the roll-out and development of AI-technologies in the internal market can be significantly improved by preventing fragmentation and increasing legal certainty through harmonised measures at EU level, compared to the baseline scenario involving possible adaptations of liability rules at national level. 104 The economic study concluded – as a conservative estimate – that targeted harmonisation measures on civil liability for AI would have a positive impact of 5 to 7 % on the production value of relevant cross-border trade as compared to the baseline scenario. 105 This added value would be generated notably through reduced fragmentation and increased legal certainty regarding stakeholders’ liability exposure. This would lower stakeholders’ legal information/representation, internal risk management and compliance costs, facilitate financial planning as well as risk estimates for insurance purposes, and enable companies – in particular SMEs – to explore new markets across borders. 106 Moreover, it is only through EU action that the desired effect of promoting consumer trust in AI-enabled products and services by preventing liability gaps linked to the specific characteristics of AI can be achieved consistently across the internal market. Harmonised adaptations of existing liability rules are needed to prevent AI-induced liability gaps across all MS, ensuring a consistent (minimum) level of protection for all victims (citizens and companies) and consistent incentives to prevent harm and ensure accountability.

Given the long-standing and entrenched divergences between national tort law traditions, any attempt of a general broad stroke harmonisation of tort law would meet with strong subsidiarity concerns from MS. The same is not the case for a targeted initiative addressing AI specific challenges at EU level. The pivotal role of AI as a key enabling technology distinguishes it from other products and services. This is why the Commission and MS have jointly committed to align their AI policy to remove fragmentation and address challenges globally 107 , recognising the need for coordinated measures to create the right conditions for the rollout of AI in the internal market. By putting forward targeted and harmonised adaptations of certain aspects of existing national liability rules, this initiative presents a clear added value, without putting into question MS’ traditional civil liability frameworks.

(c) Stakeholder views

A public consultation conducted in 2017 by the EP showed that 90 % of individual stakeholders consider it necessary to regulate developments in both robotics and AI, and 96 % of those favouring a regulatory approach preferred action at EU or international level rather than action at MS level. 108  A statistically representative Eurobarometer of 2019 confirmed that half of EU citizens (50 %) agreed that public policy intervention is needed, whereas only 7 % considered that no specific action is needed, and 16 % considered that industry providers of AI can deal with these issues themselves. 109  

These results were clearly confirmed by the outcome of the public consultation on the AI White Paper in 2020. 63 % of respondents were in favour of adapting national liability rules for all or for specific AI applications to better ensure proper compensation and a fair allocation of liability. Citizens were particularly supportive, with 72% of respondents calling for some adaptations. While the overall support among companies was 45%, it was higher among SMEs with 60% of respondents. This confirms that, as explained before, the effect of the identified legal uncertainty and fragmentation is disproportionately larger on SMEs.

The need for EU action was again confirmed by the results of the public consultation done in the context of this impact assessment. When asked to rank PO by preference, the baseline option ‘No EU action on liability for AI’ had clearly the lowest average preference score of all responses. Moreover, ca.  60 % of respondents think that EU action to ease the victim’s burden of proof is necessary and justified, while only 25 % disagree with this.

The public consultation showed particularly strong support for EU action amongst EU citizens, consumer organisations and NGOs; almost all of them chose ‘No EU action’ as their least preferred PO. Although business stakeholders tend to be more sceptical about EU action on liability for AI, there was support also from the business side: The share of responding companies/business organisations in favour of EU action to ease the victim’s burden of proof was the same as of those opposing it (ca.40 % respectively). SMEs are slightly more favourable towards EU action than other types of business stakeholders: 4 of the responding associations representing SME interests supported EU action to ease victims’ burden of proof, while 3 did not; all responding individual SMEs were in favour of EU action to address difficulties of proof linked with AI.

The comparably less pronounced support from businesses, in relation with all the other categories of stakeholders, is likely to have two main reasons.

Firstly, experience seems to indicate that business stakeholders, when contributing to public consultations of the Commission, which is considering a legislative intervention, usually tend to do so from the perspective of a future subject of obligations imposed by EU law, focusing on potential burdens, and not from the perspective of a potential beneficiary. Accordingly, in the case of the public consultation on AI liability, businesses mainly may have seen themselves as liable person/entity, and not as a potential victim of damage. The experiences in the context of the economic study, where business stakeholders provided input on the policy options from the perspective of potentially liable parties and not of potential victims of harm, seem to support such a viewpoint. In line with the above-mentioned differences in results when looking at views of businesses overall and SMEs specifically, the latter could however likely to be more able to view themselves in the victim’s position. This may be one major reason why respondents of this category had somewhat different views than large companies or business associations.

Secondly, businesses may tend to focus mostly on how the legislation would apply to them directly and not to consider the broader effects of policy measures, from which they would benefit indirectly. Businesses might tend to disregard the macroeconomic effects of building trust in a technology, while they are more aware of the effects of the lack of it. Therefore, most businesses replying to the consultation seemed to compare a situation of no intervention to a situation of regulation at EU level:

When considering no legislative intervention, on the one hand significant shares of business stakeholders acknowledged a link between a lack of adaptation of national liability rules applicable to AI and legal uncertainty and legal fragmentation. They also expect to incur costs due to these problems, especially when engaging in cross-border trade. On the other hand, many businesses may likely tend to associate the lack of adaptation of national liability rules with an indeed high probability that even if companies are responsible for harm caused by AI-enabled products and services they would not have to compensate the harm at all because potential victims would not be able to prove their otherwise justified compensation claim.

By contrast, when considering a situation of regulation at EU level, a large share of business stakeholders may focus more on the fact that harmonised measures to prevent AI-specific compensation gaps would increase the chances of success of such a liability claim against them. These considerations might be given more weight than the fact that the AI liability initiative aims to help also businesses that suffered harm to claim compensation, or the potential benefits in terms of increased legal certainty and reduced fragmentation.

Authorities from only 5 MS participated in the public consultation, expressing split views on the AI liability initiative. A dedicated workshop was subsequently organised to get a broader picture of Member States’ opinions. However this aim was not achieved, Member States were in their very large majority not in a position to confirm either clear support or opposition. Member States seemed to adopt more a ‘wait-and-see’ stance. A reason may be that they may expect the Commission to put forward a legislative proposal and may not have wished to pre-empt their position during the legislative negotiations.

4.Objectives: What is to be achieved?

4.1. General objectives

In line with the overall Commission strategy, the primary general objective of this initiative is to contribute to the roll-out of AI in order to reap its benefits as a set of enabling technologies for the growth of the digital economy and its societal advantages. The contribution of this initiative to this objective is twofold.

Firstly, it would improve the conditions for the functioning of the internal market: reducing existing obstacles and preventing the emergence of new obstacles to the cross-border trade in AI-enabled products and services. This would have beneficial effects for the European economy and the competitiveness of the European AI-sector.

Secondly, in line with the Commission human-centric approach and the White Paper on AI, this initiative contributes to the ‘ecosystem of trust’ in AI by ensuring that persons having suffered harm caused with the involvement of AI systems enjoy the same level of protection as persons having suffered harm caused by other technologies and that efficient redress mechanisms fairly distribute loss. 110  Increased trust will lead to increased uptake of AI.

4.2. Specific objectives

(a) SO1: Increase legal certainty about the liability risk exposure of business activities involving AI

- Description: This initiative aims at reducing legal uncertainty regarding companies’ liability risk exposure. 111  Legal certainty is a key concern for investors. 112  

By clarifying the conditions for assessing and quantifying liability risks, the initiative seeks to create investment stability. By reducing legal information/representation, risk management and litigation costs, companies should have stable and favourable conditions to produce, market and operate AI-enabled products and services. This initiative thus promotes the roll-out of AI-enabled products and services.

This SO applies to all AI-related business but is particularly relevant for cross-border operations, where the negative effects of legal uncertainty are multiplied by the number of MS into which AI-enabled products and services are exported to or are operating in.

- Stakeholders targeted: Legal certainty is meant to benefit all companies active in AI. Increasing legal certainty would be beneficial to a higher degree for start-ups and other SMEs, which perceive uncertainty to a greater extent and are disproportionately affected by the associated costs and burden. 113  

By reducing legal uncertainty, the initiative aims also to promote a wide range of offers in the relevant insurance market and enable insurance companies to reduce risk premiums. Those taking up AI will be able to insure their liability risk, with the effect of limiting liability exposure to the insurance premiums. Such insurance decreases possible liability costs very significantly. It also makes them predictable and therefore manageable. This benefits in particular start-ups with lower capital cover, because a liability risk is more likely to be an existential threat for such companies than for large companies. 114  Insurance coverage is also advantageous for victims, because they benefit from faster and easier compensation, even in the event that the liable party is insolvent.

(b) SO2: Prevent the emergence of fragmented AI-specific rules across the internal market

- Description: The initiative aims to prevent the emergence of different national rules adapting in different ways national liability regimes to the specificities of AI. This would exacerbate the already existing legal heterogeneity and create more fragmentation. It would also slow down the roll-out of AI as a set of enabling technologies. Legal fragmentation adds to the burden linked to legal uncertainty when operating across national borders. The second SO is thus intertwined with the first. Preventing legal fragmentation in form of diverging adaptations of liability rules at national level alleviates those types of burden that apply specifically when operating across borders. For instance, the initiative would reduce the need for companies to adapt their business models and the technological parameters of AI-enabled products and services in light of diverging proof-related liability rules.

- Stakeholders targeted: Attaining this SO would benefit primarily companies, whose activities are linked to the development or use of AI, that are already operating, or are considering to expand, across borders. As with the first SO, start-ups and other SMEs are affected to a markedly higher degree than large companies, given that the former are much more likely to be deterred by legal fragmentation from exploring new geographical markets. This SO covers also insurance companies. More consistent extra-contractual liability rules would facilitate it for them to offer competitive coverage of AI-related activities in several MS, notably by enabling a consistent risk-assessment and insurance product design.

(c) SO3: Prevent a lack of compensation in cases involving AI to promote societal trust and the uptake of AI-enabled products and services

- Description: By preventing AI-specific liability gaps, the initiative promotes societal acceptance and consumer trust, both in AI as a new technology and in the justice system. This SO forms part of the EU’s human-centric approach to create an ‘ecosystem of trust’.

- Stakeholders targeted: This initiative aims at increasing societal trust in AI-technologies 115 and access to an effective justice system. It ensures that civil liability is well functioning, adapted to the specificities of AI, and justified claims for compensation of harm are successful. It will reach this aim by facilitating claims where the specific characteristics of AI (see 2.2.) would otherwise make it difficult or prohibitively costly for victims to meet the burden of proof.

Increasing societal trust would benefit all companies in the AI-value chain, because strengthening citizens’ trust will contribute to a faster uptake of AI, and create a competitive advantage for European AI companies. Due to the incentive effect of liability rules, preventing liability gaps would also indirectly benefit all citizens through an increased level of protection of health and safety (Article 114(3) TFEU) and the obviation of sources of health risks (Article 168(1) TFEU).

(d) Consistency with other Commission proposals concerning AI and the PLD revision 116

Not only is this proposal an important element of the overall Commission strategy to promote the roll-out of AI and its human-centric approach. It will also incentivise compliance with safety requirements and requirements designed to safeguard fundamental rights, applicable to AI-enabled products and services. 117  Beside its other main objective to achieve fair compensation, liability law acts as deterrent to unwanted behaviours, in this case, a non-compliance with safety and fundamental rights requirements (see 1.3.). Potentially liable persons will be more diligent in observing rules meant to prevent harm by AI systems. This comes from the knowledge that, in addition to possible administrative sanctions (e.g. fines, withdrawal of products, and suspension of licences), they face the risk of having to pay compensation for a damage caused.

The general and SO of this initiative are also consistent with those of the general PLD revision (PLD impact assessment section 4). The two initiatives are designed to complement one another. In their combination, they will create a clear and consistent legal framework ensuring effective protection of victims of AI. Given the PLD’s horizontal, technology-neutral approach, the PLD revision is not aimed at AI in particular but covers all kinds of products. This initiative addresses AI-specific challenges. 118 The PLD revision pursues also additional objectives, such as making the PLD fit for the circular economy.

(e) Contribution to achieving the Sustainable Development Goals

By promoting the development and rollout of AI, this initiative indirectly supports several targets across all the Sustainable Development Goals (SDGs). TA more direct contribution of this initiative to the SDGs is towards Goal 16, especially target 16.3. This initiative will ensure that victims of harm caused by AI-enabled products/ services have effective access to the evidence necessary to build a case in court, so as to enjoy the same chances of a successful claim as victims of harm caused by other technologies. By ensuring to the former the same level of protection as to the latter, this initiative promotes the goal of ‘equal access to justice for all’.

5.What are the available policy options?

5.1. What is the baseline from which options are assessed?

(a)    Description

The baseline builds on the existing legal framework outlined under 1.2., which consists primarily of MS’ fault-based and strict liability rules as well as the PLD. It is also assumed that the proposed AI Act 119 , the Machinery Products Regulation (‘MPR’) and the General Product Safety Regulation will be adopted. The expected adaptation of national liability rules by a number of MS, likely leading to further fragmentation, is taken into consideration when assessing the impacts of the identified problems under the baseline. A time horizon of 10 years is envisaged for the baseline and for the assessment of the PO against that baseline. 120  

 (b) Impacts of problems under the baseline scenario

- Market barriers linked to legal uncertainty and fragmentation (problems 1 and 2): 122  It is likely that at least some MS will adapt their respective liability rules under the baseline scenario. Depending on their respective approaches, this may increase legal certainty to a varying extent. However, legal uncertainty would not be addressed in a systematic and coherent manner across the internal market. In those MS that will not adopt general legislative measures, only ad hoc judicial solutions will emerge, based on a case-by-case interpretation of existing liability rules. Affected businesses, in particular SMEs, are therefore likely to continue facing most of the costs of legal uncertainty described under 2.6. In addition, the national legal or case-law solutions reacting to AI specific challenges would increase the level of legal fragmentation, which would lead to higher costs for companies operating cross-border.

With the expected roll-out of increasingly autonomous, opaque and complex AI-systems over the coming years, AI will be implemented in ever more sectors of the economy and society. It is therefore highly likely that damages and ensuing liability claims involving AI-systems will become more frequent, and the economic relevance of the liability-related problems will increase under the baseline scenario. It was estimated that the total value of the EU27 market shares affected by those problems will increase almost seven- to ten-fold between 2021 and 2025 (from between EUR 1.739 billion to EUR 4.973 billion to between EUR 10.204 billion and EUR 21.342 billion. 123 The market shares affected by legal uncertainty and fragmentation were determined by a sector-by-sector analysis in the economic study, taking into account the current and forecast take-up rate of AI-enabled technologies, the percentage of companies perceiving liability as a barrier to AI adoption and the ability of the AI applications deployed in the respective sector (now or in the foreseeable future) to cause harm and thus give rise to liability claims. 124 For all of these factors, the study made reasoned assumptions based on data from the aforementioned representative company survey on AI, relevant individual stakeholder feedback as well as a detailed market analysis. The size of the market shares affected by legal uncertainty and fragmentation provides an indication of the order of magnitude of the economic relevance of the problems to be addressed by this initiative. This indicator is not intended to provide a precise quantification of the economic impact of the legal uncertainty and fragmentation. Such quantification proved impossible due to a lack of data, as the AI-enabled products and services affected by the initiative are in the early stages of rollout or not yet on the market. It is also acknowledged that the factors and assumptions supporting these estimates are to a certain extent uncertain, as they reflect uncertain projections on future technological and market development. In particular, the assumption that the size of the market shares affected by legal uncertainty and fragmentation will increase proportionally to the increase in the overall size of the AI market depends on a number of yet unknown factors, such as the extent to which revised safety rules will decrease the instances of harm caused by AI and possible advances in the field of explainable AI. Taking into account that the increasing degree of complexity and behavioural autonomy of AI systems is likely to exacerbate the problems to be addressed by the AI liability initiative, and that MS will over time become increasingly likely to adapt their liability rules unilaterally in the absence of EU action, the assumption of a proportional increase of the market shares affected by legal uncertainty and fragmentation appears conservative. In any case, a maximum of robustness is ensured by a allowing for a wide bracket between the high and low estimates, which can accommodate differing market definitions and future developments.

- Lack of compensation for victims leading to a lack of societal trust and hampering take-up by consumers (problem 3): In those MS that adapt their liability frameworks to the challenges of AI, claims for compensation could become similarly effective as in cases where no AI-technologies are involved. However, in those MS which will refrain from adapting liability rules reacting to AI specific challenges during the coming years it will often be prohibitively difficult or even impossible for victims to claim compensation.

Over the course of the baseline period, the increasing prevalence of AI-systems in more and more sectors of the economy and society will mean more accidents and more liability claims that are likely to be attract significant media attention. This will drive consumers’ awareness of liability-related issues. As confirmed by dedicated behavioural research, persisting – and increasingly mediatised – safety concerns will in turn also contribute to a lack of societal trust in AI. 125 Thus, liability-specific issues, if not addressed consistently at EU level, are expected to have a significant negative effect on societal trust and the uptake of AI. 126  That is because lack of trust means less willingness to engage with the technology. Delayed or reduced uptake of AI-technologies in the internal market will affect all companies in the AI-value chain. It will also lead to reduced or delayed benefits of AI for the society.

The practical consequence of this situation is both unfair and economically inefficient. On the one hand, certain risk-takers that benefit from integrating AI into their activities will be able to externalise to the victims the costs linked with the use of AI. On the other hand, victims – be they natural persons or companies – will be burdened with costs that they have no means to prevent. The liability framework would thus fail to fulfil its functions of preventing unwanted behaviours and ensuring an efficient cost allocation of risks.

Persistent liability gaps are also likely to frustrate, to some extent, the White Paper objective of ensuring the safety of AI for the benefit of all EU citizens, as liability rules will not incentivise potentially liable persons to prevent accidents.

(c) The relevance of liability-specific problems given the AI Act, MPR and General Product Safety Regulation

The legislative proposals (in particular the AI Act and MPR) were also adopted as Commission’s follow-up to the AI White Paper to address specific problems of AI and they pursue the same overall objective as the present initiative. 127  

These proposals introduce requirements to be met before AI systems placed on the market, to reduce safety risks and breaches of fundamental rights. They also provide for monitoring of how these systems operate in practice and reporting of incidents.. These provisions reflect that there is no absolute safety, only a level of accepted risk. The impact assessments accompanying the proposed AI Act, MPR and General Product Safety Regulation indeed conclude that these initiatives will reduce accident rates, while quantified estimates of the expected safety gains regarding AI-enabled products or services in particular are not available. It is therefore assumed, albeit only in qualitative terms, that safety levels will increase during the baseline period, an assumption that has been incorporated in this impact assessment.

Hence, it is certain that these proposals, while reducing the risk for accidents, will not exclude harm caused by AI systems. The AI liability initiative will deal only with those accidents happening once the AI Act and other relevant safety legislation are adopted and applied.

Once adopted, the requirements that market players (e.g. professional users) have under the AI Act could inform the standard of care that courts will have to assess in liability claims against them. Similarly, the information to be documented/logged pursuant to the AI Act could be used in liability claims, to the extent victims could have access to such evidence under various national laws. However, in line with the staged approach of the Commission’s AI policy, and the complementary nature of safety and liability rules, the provisions of the AI Act and the MPR are not designed to address the problem drivers and problems tackled by this initiative. Their provisions are not designed to clarify how liability rules should apply to AI and to facilitate the individual compensation of victims of harm. This means that, under the baseline scenario, the liability situation will remain subject to the same challenges as explained above. The Commission aims to address these challenges and achieve synergies between these proposals and the envisaged civil liability related measures (see s.1.3. and 2.8 and Annex 7).

5.2. Description of the policy options

Legal instrument for implementing the PO: Each of the PO could be implemented either by a binding legislative instrument or by a non-binding recommendation addressed to MS to implement the preferred measures in their national law. The approach for selecting the suitable instrument is set out in Chapter 6. 128

(a)    PO1: Easing the burden of proof for AI-related claims

Description: The problems are primarily linked to the difficulty of the victim to prove the necessary facts due to the specific characteristics of certain AI systems. This PO would help the victim to handle its burden of proof in a way that justified liability claims can be successful.

PO1 would involve three complementary measures. Two of these – (i) and (ii) – are closely aligned with the approach and duties set by the AI Act. Definitions of the AI Act (AI-system, provider, user) would be taken over to ensure consistency. None of these two measures would involve the introduction of any new risk categories; the risk-categorisation of the AI Act would be mirrored. They would be fully compatible with the confidentiality safeguards included in the AI Act. Measure (iii) would not involve any differentiation according to the risk profile of AI systems. None of the measures of PO1 would provide for any substantive obligations regarding the development or use of AI systems.

PO1 would not define either what constitutes in national law ‘fault’ or ‘causality’. MS would be able to fit provisions into their national liability regimes without disrupting their legal traditions.

(I)The disclosure, for liability claims, of information that the user or provider of a high-risk AI has to record or document pursuant to the AI Act.

Obtaining information recorded or documented pursuant to the AI Act 129 could help the victim to make a successful liability claim. 130  For instance, it could help to provide evidence for demonstrating fault as one of the conditions of liability to be proven by the victim. This could be done for example by proving that the liable person did not comply with their obligations under the AI Act.

According to this measure, the victim could request from the possibly liable party the disclosure of such information concerning the victim’s claim. The competent national court could order that such disclosure would be subject to stringent safeguards to ensure proportionality and protect the legitimate interests of all parties concerned, for instance confidential information, intellectual property rights and trade secrets.

If the possibly liable person does not disclose the information, the victim would benefit from a presumption that the facts, which the withheld information might have proven actually, did occur. Such presumption would facilitate the burden of proof of the victim.

The liable person would have the opportunity to rebut the presumption. This means that the liable person would have the opportunity to prove that the presumed facts did in reality not occur. That person could also avoid liability by refuting any of the liability conditions set by the applicable liability rules, for instance demonstrate that the relevant AI system was carefully selected and used, taking into account the requirements of the AI Act.

This measure would not only help the victim. Providers and users of AI systems would be given an additional incentive to comply with the documentation/logging requirements of the AI Act. If they had complied, they could disclose the information (subject to necessary confidentiality safeguards) and would be able to demonstrate that they acted in accordance with legal requirements. This additional incentive would support the compliance with the AI Act. Thereby this measure would achieve synergies with the AI Act.

(II)If the victim shows that the liable person did not comply with AI Act requirements designed to prevent damage, the second measure could establish a rebuttable presumption that this non-compliance caused the damage. 

This measure would further facilitate the exercise of the burden of proof as regards the causality link between fault and damage for the victim. 131  The victim could demonstrate that the liable person did not comply with AI Act requirements. In that case the victim could claim and the court would presume for the purposes of the victim’s liability claim that such non-compliance would have caused the harm. Again, such presumption would be rebuttable. This means that the possibly liable person could avoid liability by proving that their non-compliance with the AI Act requirements did not cause the damage.

In line with the AI Act, this measure would be primarily relevant for claims against providers and users of high-risk AI-systems. It would give them an additional incentive to comply with the AI Act, because in case of non-compliance it would be presumed that their behaviour caused the damage.

However, this measure is not be sufficient to ensure that victims always have the same level of protection when suffering harm caused by AI:

-Firstly, the information needed by the victim to prove non-compliance may for instance not be recorded. In such cases, the victim may not be able to establish the liable party’s non-compliance with the AI Act (and thus to trigger the presumption). For example, the victim may not be able to know what kind of input data the AI system was exposed to or what circumstances of use the AI system was designed for.

-Secondly, the presumption can apply only if a ‘high risk’ AI-system as defined in the AI Act is at stake. It would therefore not apply to AI systems, like for instance autonomous drones, which are not covered by the AI Act requirements, but can cause harm.

-Thirdly, the presumption would apply only if the requirement that the liable party did not comply with, was intended to prevent harm of a type suffered by the victim. For example, if the victim seeks compensation for damage caused by discrimination, but can only establish non-compliance with a requirement meant to prevent physical harm, this would not trigger the presumption.

(III)Targeted alleviation of the exercise of the burden of proof, so that the victim does not have to explain the inner workings of the AI. 

As explained under 2.2., the specific characteristics of the AI-system could make the victim’s burden of proof prohibitively difficult or even impossible to meet. There will be cases when the victim will not be able to identify or explain how something that the liable person did (or failed to do) was ‘translated’ by the AI into the output that caused the damage.

The existing legal framework offers a limited set of alternative tools to help victims overcome these proof-related difficulties:

-Rebuttable presumptions: If the victim meets a reduced burden of proof, e.g. by demonstrating the plausibility or likelihood of certain facts, it is presumed that those facts occurred. In order to avoid having to pay compensation, it is then for the liable party to demonstrate that these facts did in reality not occur or that other facts, for which the liable party is not responsible, occurred. This tool leaves the basic distribution of the burden of proof intact, but makes it easier for the victim to discharge that burden.

-Shifts/reversals of the burden of proof: This tool departs from the general principle that the party invoking a legal provision has to prove that the conditions for the application of that provision are fulfilled. The victim hence no longer bears the burden of proof, but it is for the liable person to prove that one or all conditions of liability are not fulfilled. Such a solution was proposed by the European Parliament for fault-based liability. According to that resolution, the operator of an AI system would by default be liable for harm or damage caused by that system, unless they can prove that the harm or damage was caused without his or her fault.

-Irrebuttable presumptions: This tool goes even further than a reversal of the burden of proof, as it posits certain conditions or facts as given, without allowing for the possibility to proof that this presumption is wrong.

All of these possible approaches have been analysed when designing measure (iii), giving particular consideration to the suggestions put forward by the European Parliament. For the following reasons, only the tool of a rebuttable presumption has been retained in the PO for detailed assessment:

-MS’ civil laws are based on the principle that the victim has to prove all the facts necessary to support their claim. A rebuttable presumption would leave this fundamental principle in place. This approach would make it easier for MS to integrate the harmonised AI-specific adaptations of liability rules into their respective legal systems without friction. It also preserves a maximum of consistency between the rules applicable in cases involving AI and the general rules applying in other cases.

-The AI liability initiative seeks to help victims meet their burden in the most targeted, i.e. proportionate manner possible. A rebuttable presumption is the least interfering way of alleviating the burden of proof.

-A reversal of the burden of proof regarding fault and causation, as suggested by the European Parliament, or, to an even greater extent, an irrebuttable presumption represent a more far-reaching exception from MS’ general civil liability rules, which puts into doubt their political feasibility. 

-Such approaches would create a situation where victims of damage caused with the involvement of AI would systematically be in a more favourable situation than victims of damage caused by other technologies. They would thus overshoot the Commission’s policy objective defined in the AI White Paper, and risk dis-incentivising the uptake of AI.

-The public consultation showed strong opposition from business stakeholders against a shift of the burden of proof: 63 % of these respondents (54 out 86) disagreed with such an option, and only 24,4 % (21 out of 86) agreed. Taking into account the initiative’s objective to promote the rollout of AI, this result militates in favour of alleviating the burden of proof through less interfering measures, that is to say in favour of a rebuttable presumption instead of a reversal of the burden of proof or a irrebuttable presumption.     

The rebuttable presumption to implement measure (iii) is a procedural tool, to help in court proceedings, that would intervene only in the context of existing fault-based liability claims. 133  If the victim makes it sufficiently plausible that an action or omission of the potentially liable party led the AI system to arrive at the damaging output, the court could presume that this was indeed the case. The potentially liable party still has the possibility to rebut such presumption. The measure would not involve the creation of new risk categories, substantive obligations or liabilities.

This measure would not be relevant where AI systems are used to provide advice or information to human decision-makers (e.g. medical analysis AI informing the diagnosis and treatment decisions of human physicians). In such cases the AI system is not interposed in the causal chain between the relevant human action and the damage. It will thus not be necessary for the victim to establish what triggered a specific output of the AI system.

Scope of PO1: Which AI systems would be concerned?

Measures (i), (ii) and (iii) would apply where an AI system as defined in the AI Act has caused harm. While the specific challenges of AI to liability rules are driven by the peculiar characteristics of certain AI systems, it is appropriate to define AI systems by reference to the AI Act’s broad concept of AI rather than limiting its application to a subset of AI systems. As measures (i) and (ii) are directly linked to the AI Act, their scope must be aligned to it, in any case, so they would apply to high-risk AI systems, as defined by the AI Act. In addition, victims of harm should not have to demonstrate that an AI system has specific characteristics or belongs to a specific category in order to benefit from the envisaged measures. The application of PO1, including measure (iii), should therefore not hinge on such criteria. Otherwise the objective of preventing a lack of compensation and ensuring legal certainty would be in doubt.

However, the impact of the measures will vary depending on the specific characteristics of AI systems. Deterministic, transparent and relatively simple AI systems entail fewer difficulties of proof so the possibility of rebutting the presumption will be in practice particularly relevant for those AI systems. The beneficial effect of the rebuttable presumption alleviating the victim’s burden of proof will be thus particularly relevant for those AI systems which show the specific characteristics of opacity, autonomous behaviour, complexity or limited predictability.

Addressing legal uncertainty and legal fragmentation and related internal market obstacles (2.2., 2.3., 2.6.): PO1 would provide uniform rules clarifying how national burden of proof rules would be applied to the specific challenges of AI. This would avoid that national courts need to employ ad-hoc solutions in each case before them, to achieve equitable solutions for adequate victim compensation. It would also prevent that MS see a need to adapt, based on their own different approaches, the relevant burden of proof rules, to react to the characteristics of AI challenging the application of those rules.

PO1 will provide the legal certainty needed by businesses – in particular SMEs – to foresee possible liability risks and insure themselves against them. Accordingly, they will have investment stability to roll-out AI-enabled products and services in the internal market. Clear conditions for meeting the burden of proof would also facilitate ex-ante risk-assessments by insurers offering coverage of civil liability for damage caused by AI.

Addressing the possible lack of compensation for victims leading to a lack of societal trust and hampering take-up by consumers (2.5., 2.6.): Measures designed to alleviate the exercise of the burden of proof by the victim allow victims to have successful liability claims obtaining compensation in cases of damage created by AI. By avoiding that victims are less protected due to the involvement of AI than victims of traditional technologies, PO1 contributes to the creation of an ecosystem of trust for AI. It thereby promotes the uptake of AI-enabled products and services in the EU 134 . All companies in the AI value chain would benefit from increased demand.

Incentive effects: Finally, PO1 would have the ancillary purpose of incentivising compliance with the AI Act and damage prevention. It will thus increase the level of protection for all persons affected by the use of AI.

How does this PO interact with the measures envisaged for the PLD revision?

In line with the AI White Paper objective, all PO need to ensure that victims having suffered harm caused by AI systems enjoy the same level of protection as those having suffered harm caused by other technologies. PO1 will therefore introduce measures to ensure that victims of AI can have effective claims based on fault, despite the specific characteristics of AI that make the burden of proof very difficult.

In claims against the producer, the PLD revision will ease the burden of proof for damages caused by defects in complex products, including AI-enabled products, by rules on when producers are obliged to disclose technical information and rebuttable presumptions. Moreover, the development risk defence, which exempts producers from liability when a product’s defective nature was not scientifically discoverable at the moment it was put into circulation, would be adapted to take account of the dynamic nature of products with digital elements, such as AI-enabled robots. If the defective nature becomes scientifically discoverable while the producer retains control of the product’s safety features (e.g. through software updates), the defence would not be available.

PO1 uses very similar tools to ensure that a fault based claim for compensation for harm induced by AI is as effective as for harm induced by other technologies. Compared to the PLD, PO1 covers additional forms of damages, additional victims and additional liable persons as well as different sources of damage. Thus, alleviations of burden of proof are complementary between the two initiatives, as the following examples illustrate:

-the producer’s obligation to give the victim access to information for claims under the PLD is complemented by the disclosure of information, based on this initiative about liability for AI, in claims based on the user’s fault or in claims against the producer for damage suffered by business, 

-presumptions of defectiveness and causality under the PLD are complemented by a targeted presumption of causality, for the purposes of other civil liability claims under this initiative, when the liable person does not comply with the AI Act.

In their combination these measures will ensure that victims can use effectively all available routes to compensation. Thereby they enjoy the same level of protection compared to damage created by other technologies. They will also provide the right incentives for all entities involved in the ‘life cycle’ of AI-enabled products and services to follow obligations meant to prevent harm from happening. This way, the liability risks are distributed over the AI value chain to the appropriate liable person.

(b)    PO2: PO1 + strict liability for AI use-cases with a special risk-profile

PO2 combines the alleviations of the burden of proof envisaged under PO1 with a harmonised risk-based strict liability regime applicable to professional operators/users of certain highly autonomous AI-enabled products or services with a special risk profile.

In order to ensure, in line with the AI White Paper objective, that victims have equally effective compensation paths for damages caused by AI and other technologies, all pillars of liability need to be considered. All MS currently have, in addition to fault-based and PLD, also strict liability regimes. While they vary in scope and conditions, they usually cover risk of harm to high-ranking legal interests such as life, health, and property. These regimes assign liability to the person benefitting from the source of risk (usually owners, users, operators). In strict liability claims, the victim only has to prove that the risk ‘materialised’. As a consequence, the liable person determined by law has to compensate the damage. Thus, strict liability is an easier path to compensation for victims.  135

Following a targeted and risk-based approach, PO2 is based on the consideration that the operator/user is the one that decides on and benefits from the use of AI-enabled products or provision of AI-enabled services in practice. 136 The operation of certain AI-enabled products and services creates a risk for the general public and high-ranking legal interests. For instance, innocent bystanders, who do not choose to submit themselves to such risks and are mostly even unaware of them, should in such cases have an uncomplicated way of obtaining compensation for the harm suffered. Strict liability could be that solution.

Such harmonising measure would be limited to strict liability of those persons operating AI-enabled products and services as part of their professional activity. This personal scope is justified by the fact that it is more likely that professionals would be able to control the operation of such products and services as part of their business activities, while consumers typically have little or no influence on how highly autonomous AI systems function and therefore no means to control the specific risk presented by such AI systems.. Moreover, this scope is consistent with the approach of the AI Act and existing and future safety-relevant legislation, which impose obligations on professionals only.

Limiting the personal scope to professional users/operators is justified by the fact that consumers typically have no influence on how highly autonomous AI systems function and therefore no means to control the specific risk presented by such AI systems. For this reason, it is doubtful whether the operation of AI-enabled products with a potential to cause significant harm to legal interests of high value will at all be entrusted to consumers in the future, in particular in scenarios where such products could cause harm to unwitting third parties. Indeed, the AI Act shows that professional users have a key role in controlling the risks linked to the operation of AI systems. Should cases emerge where consumers play the role of a user/operator of AI technologies with a ‘strict liability profile’, these cases are highly likely to be already covered by strict liability in almost all MS, as exemplified by the use-case of autonomous vehicles. The risk that the harmonisation of strict liability for professional users/operators would create situations where the level of protection varies arbitrarily, from the perspective of the victim, compared to cases where the same type of AI system was used by a consumer, is therefore very small.

Similar to a common approach in MS 137 , this limited strict liability regime could be coupled with an obligation for liable persons to insure themselves. This corresponds e.g. to the present situation for car accidents, as harmonised by the Motor Insurance Directive. In addition to the measures outlined for PO1 (measures i-iii), PO2 would thus comprise:

(I)A harmonised strict liability regime applicable to the person exercising control over the operation of the relevant highly autonomous AI technology (operator/user) in the course of their professional activity. The professional in whose name the relevant product is registered or who is registered as service provider could be presumed to be the operator. Like the approach of the AI Act, the relevant products and services could be defined by a list in a technical Annex, complemented by the possibility for the Commission to update that list by delegated acts to technological and market developments. The AI use-cases included in the scope of the strict liability regime would be identified on the basis of criteria consistent with MS’ legal traditions, such as the ability to cause frequent or severe harm, in particular to high-ranking legal interests like life, health or property, and the likelihood that the public at large or at least unwitting third parties are exposed to such risks. These criteria could, for instance, be met by technologies like autonomous drones or vehicles. In contrast, strict liability would likely not be appropriate, for example, for merely stationary robots operating exclusively in confined environments and presenting risks to only a narrow, pre-defined range of people.  

In addition, only the grounds for excluding or reducing strict liability would be harmonised (e.g. force majeure and own fault), while other modalities would be left to national law.

(II)As a cumulative sub-option, an insurance obligation for the strictly liable persons, combined with a right for the victim to make a direct claim against their insurance undertaking. The insurer who paid compensation would be entitled to claim it back from any other party that could be liable for the damage under other compensation paths (e.g. PLD or fault-based).

Because of the high interests at stake and the victim’s particularly vulnerable position, strict liability ensures that compensation is promptly and easily granted to the victim. However, this does not imply that somebody else might not be at fault for the damage or, in the case of products, that the AI-enabled product was not defective. Therefore, it is relevant for the strictly liable person and insurance of that person to have effective liability claims for eventually recovering the compensation paid to the victim from the person ultimately responsible for the damage.

The envisaged harmonised measures would not create additional layers of regulation. Where MS already provide for strict liability for the cases covered by the harmonising measure, mainly because that type of activity or goods are already subject to strict liability (e.g. cars in almost all MS), those MS would have already transposed the measures under PO2. When EU law already provides for respective insurance requirements (e.g. mandatory insurance for motor vehicles or for drones above 20 kg), that requirement would continue to apply. It would neither be replaced nor superimposed by an additional insurance requirement for certain AI systems.

Approach to addressing legal uncertainty and legal fragmentation and related internal market obstacles (2.3. + 2.4.): By specifying the AI-enabled technologies subject to strict operator liability at EU level, PO2 aims to increase legal certainty and predictability of liability risks for potentially liable parties. It also prevents the emergence of divergent AI-specific strict liability regimes at national level. The approach underpinning PO2 relies on a similar logic, a targeted and risk-based approach, as PO1. However, it goes further into MS’ liability systems and into the risk distribution between stakeholders because of the special risk profile of the AI technologies subject to strict liability. The possibility for operators to cover their liability risk with (mandatory or voluntary) insurance, thus reducing their risk exposure to the annual premium, is an important component of this PO.

Approach to addressing the possible lack of compensation for victims leading to a lack of societal trust and hampering take-up by consumers (2.5.): Strict liability of the operator/user addresses the proof-related challenges posed by AI. The victim only has to prove that a certain risk materialised in the sphere of the liable person, instead of having to establish a misconduct. 138  The sub-option of mandatory insurance would relieve the victim of the liable party’s insolvency risk. These features of the PO would increase societal trust in AI, even when it has a specific risk profile. It would thereby promote its roll-out in the internal market.

How does this PO interact with the measures under the PLD revision? The existing liability system offers victims different avenues to seek compensation from various liable persons, under different instruments and for different types of harm (defect-based liability of producers under the PLD; national fault-based liability of any ‘wrongdoer’; national strict liability for certain specified risks). It is therefore necessary to work on all the existing avenues of compensation to ensure the same level of victim protection. These avenues attach liability to different grounds and the proposed measures will ensure coherence.

The preferred option for the PLD revision would extend strict liability to the person who produces software (including AI) or provides digital services necessary to make a product work. PO2 would harmonise strict liability of the person who operates products equipped with AI or provides services equipped with AI under particular circumstances of risk.

It can, of course, happen that the same person is linked to a damage in different capacities, and therefore fulfils the conditions of liability under different instruments. Given that this is a well-known and accepted situation under the existing national liability systems and an expression of the complementary nature of the pillars of liability, it is not expected to negatively affect innovation or uptake of AI-enabled products compared to the baseline scenario. For example, a producer of autonomous vehicles providing transport services with a fleet of such vehicles would be liable under the PLD for defects of the vehicles in their capacity as producer and under the harmonised strict liability regime for AI, in their capacity as user/operator.

(c)    PO3: PO1 + targeted review regarding strict liability and mandatory insurance measures

PO3 follows a staged approach, consisting in the implementation of PO1 in a first step, followed in a second step by a targeted review and re-assessment of the need for harmonising strict liability of the operator/user and mandatory insurance elements of PO2. Going beyond an evaluation of the already adopted measures, the targeted review would involve a dedicated mechanism designed to inform the future policy decisions on those additional elements. While the usual, backward-looking evaluation of the effectiveness, efficiency and relevance of already enacted EU rules would be done also under PO1 in line with general better regulation rules, the targeted review under PO3 takes a different, forward-looking perspective and provides for additional substantial steps to be taken by the Commission. Although PO3 would not (yet) implement all elements suggested by the European Parliament’s legislative own-initiative resolution, an EU instrument including the targeted review mechanism would nevertheless reflect the approach suggested by the Parliament, that is to say a combination of measures to ease the burden of proof for fault-based claims with a harmonised strict liability regime for certain specific AI applications. It would also demonstrate that the Commission is ready to act on the opinions expressed by non-business stakeholders in the public consultation, who showed clear support for strict liability. The targeted review provides a framework and prepares the ground for the Commission to take a policy decision specifically on the harmonisation of strict liability and mandatory insurance, at a later point in time when all the information for that decision will be available.

The review would notably cover the following points:

-the market developments regarding the rollout of products and services enabled by AI-systems with the characteristics that challenge existing (primarily fault-based) liability rules: a high degree of behavioural autonomy, opacity (complexity + lack of transparency), continuous adaptation and limited predictability;

-the risk-profile and operating environment of the products and services, in particular whether they can cause harm to high-ranking legal interests of the public at large;

-the incidence rate of accidents caused by AI systems, in particular those involving harm to high-ranking legal interests of unwitting third parties in the course of the operation of the relevant products or provision of the relevant services;

-any AI-specific liability gaps regarding damage caused by the identified products and services, despite the alleviations of the burden of proof implemented under PO1.

Aside from assessing to what extent PO1, in combination with the regulatory measures taken to prevent highly autonomous AI-enabled technologies from causing harm (e.g. the AI Act, the MPR, the General Product Safety Regulation, future measures under the ‘old approach’ safety legislation) achieves the Commission’s policy objectives effectively, the review will allow to conclude if the harmonisation of strict liability and mandatory insurance elements of PO2 are needed as well.

The staged approach would also allow the European insurance industry to acquire sufficient actuarial data on the realisation of AI-specific liability risks, which was a point consistently raised by insurers in the consultation activities. PO3 would enable the Commission to take into account how the insurance market for AI has evolved, when re-assessing the need for AI-specific mandatory insurance during the targeted review. Future regulatory developments at EU or national level could also be taken into consideration, namely on AI-enabled products and services with a relevant risk profile (e.g. autonomous vehicles and unmanned aircraft). For measures to ensure that the targeted review will be based on a sufficient evidentiary basis, see Section 9 and Annex 12 on monitoring and evaluation.

How does this PO interact with the PLD revision? At the first stage, the interplay with the PLD revision is the same as under PO1. In case of the introduction of strict liability as the outcome of the targeted review, the interplay would be as described under PO2.

5.3. Measures discarded at an early stage

(a)     Harmonisation of risk-based liability for damage caused by all AI-enabled products or services, irrespective of their risk-profile, possibly coupled with mandatory insurance

A horizontal strict liability regime applicable to users/operators of AI-applications irrespective of the risk linked to their activities would be neither proportionate nor suitable to achieve the objectives of this initiative. While such an approach may increase legal certainty and prevent fragmentation between possible future AI-specific national liability rules to the highest extent, possible benefits for companies seeking to invest in AI would be negated by the far-reaching shift of liability risks and associated compliance costs. 139 This discarded PO would entail high adaptation costs for companies. Moreover, it is uncertain whether the insurance industry would be able to cover the new strict liability risks across the broad range of AI-enabled technologies. 140  

The combined effects would likely dis-incentivise investments in such technologies, stifling innovation and slowing down the take-up of AI by European businesses, 141 having an overall negative effect on cross-border trade in AI-enabled products and services. 142  

This discarded PO would go beyond the objective to ensure that persons harmed by AI enjoy a similar level of protection as persons having suffered harm caused by other technologies. Prescribing strict liability whenever AI-systems are involved would, in many cases, treat victims of AI-induced damage better than victims of traditional technologies. 143  

(b)     Harmonisation of the types of – in particular immaterial – harm giving rise to civil liability claims when caused by AI

In its own-initiative resolution, the EP requested the Commission to analyse in depth the legal traditions in all MS and their existing national laws that grant compensation for immaterial harm, in order to evaluate if the inclusion of immaterial harm in AI-specific legislative acts is necessary and if it contradicts the existing Union legal framework or undermines the national law of the MS. The Commission’s analysis has shown that the legal situation in MS regarding the compensability of non-material harm, as well as pure economic loss, is fragmented. 144 The matter is marked by differing long-standing legal traditions. It is not an AI-specific question but rather a horizontal one.

Creating a situation whereby victims might be able to claim compensation for a certain type of harm, for instance immaterial harm, only if it was caused by AI would not be in line with the Parliament’s and the Commission’s common objective to ensure that victims have the same level of protection. Such an approach would give victims of harm caused by AI-systems a better protection compared to harm caused by other technologies.

It is therefore appropriate not to include a harmonisation of the compensable harm in the envisaged AI-specific instrument on civil liability. By ensuring the effectiveness of existing liability rules, the envisaged policy measures will nevertheless help victims to claim compensation also for other types of harm such as immaterial harm or pure economic loss, to the extent that such harm is compensable under national liability rules.

6.What are the impacts of the policy options?

The impacts of the policy options are, for the most part, economic in nature. However the initiative will also have relevant social impacts, mainly related to victim compensation and incentivising the prevention of harm. A deeper assessment of impacts was focused on the most significant ones, identified based on stakeholders’ views.  145

The quantification of data was to a large extent not possible, due to the novelty of the technology and the scarcity of products and cases of damage or compensation. Mitigation actions were taken that led to the use-case modelling of costs of insurance, macro-economic approaches to estimate impact of legal fragmentation, as well as micro-economic approach for a specific use case. The available quantified estimates (production value affected by the obstacles of legal uncertainty and fragmentation of liability rules of the use-cases, overall market value affected by these obstacles, costs of claiming compensation) were then used to extrapolate quantified estimates of the impacts of policy options on the costs linked to the burden of proof, on the EU AI market value and of the incremental change in insurance premiums. In light of the scarcity of quantitative evidence, the costs and benefits of the policy options were largely assessed using qualitative scales, while taking into account the available quantified estimates.

For the cost quantification, administrative costs are less significant under this initiative compared with other sectoral legislation because there are no specific obligations or information requirements imposed on economic operators. There will be some adjustment costs for potentially liable parties, related to liability insurance premiums. 146

6.1. PO1: Easing the burden of proof for AI-related claims

(a)Effectiveness

- Degree to which SO1 and 2 147 would be achieved: Targeted, risk-based alleviations of the burden of proof aligned with the AI Act would address the major sources of legal uncertainty identified in the problem analysis. It would thereby achieve the more general objective to improve conditions for cross-border business activities involving AI.

- Degree to which SO3 would be achieved: The envisaged alleviations of the burden of proof would effectively prevent AI-induced compensation gaps, and thus be suitable to ensure that victims suffering harm caused by AI – whether they are consumers or businesses – have the same level of protection as victims harmed by other technologies. Victims would be relieved from having to overcome the characteristic opacity of certain AI-systems to prove their claims. 148 Consequently, they would have to spend less on technical expertise and have better prospects of making a successful claim. 149  

The perceived low likelihood of compensation and the difficulty to determine who is liable count amongst the most relevant reasons for low levels of consumer trust in and societal acceptance of AI. 150 Consumers who perceive liability rules as appropriate to protect victims of harm are significantly more willing to take up such products and services. 151 A liability regime where the burden of proof has been adapted in favour of the victim ranks higher in the perception of consumers than a regime where the victim bears the full burden of proof. 152  PO1 is therefore expected to contribute – together with the already proposed adaptations of safety rules – to increasing the level of societal trust in AI-enabled products and services and consumers’ willingness to take up such products and services. 

- Indirect social impacts: By preventing liability deficits, PO1 would provide an effective incentive to prevent harm and thus drive safety-enhancing innovation. It would thus contribute indirectly to people’s overall level of safety. 153 This mechanism would apply firstly to businesses subject to specific safety requirements – in particular the user and provider under the AI Act. Secondly, by ensuring the effectiveness of general liability rules under national law, the incentive effect of PO1 could extend to any stakeholders whose actions or omissions may have contributed to the causation of damage, such as e.g. providers of labelled training or testing data. Moreover, behavioural research has shown that adapting the burden of proof in favour of the injured party makes people more likely to consider that victims receive just compensation and that the legal framework is reasonable, predictable and transparent. 154 By promoting effective access to justice, PO1 is hence likely to maintain or increase societal trust in the justice system.

(b)Efficiency (costs and benefits for affected stakeholders)

- Impacts on potentially liable parties (businesses and natural persons): The reduction of legal uncertainty and fragmentation would generate direct regulatory benefits for businesses, namely through lower legal information/representation, internal risk management and other compliance-related costs, as well as additional cross-border revenue. 155 By clarifying the kind of information and evidence potentially liable parties may be required to submit in civil proceedings, PO1 would help them to choose more efficiently between different technological options 156 , namely by favouring more transparent and explainable solutions. As start-ups and other SMEs are significantly more affected by the internal market barriers created by legal uncertainty and fragmentation (see 2.6.), this stakeholder group would also benefit to a higher degree. 157  The expected positive impacts of PO1 on societal trust in AI and consumers’ willingness to take up AI-enabled products and services, as well as the improved competitiveness of the European AI sector would directly or indirectly benefit all companies in the AI value chain. 158  

These benefits, which have been approximated in terms of increases in the AI market size by EUR 500mln to EUR 1.1bln 159 , are likely to outweigh the following adaptation (substantive compliance) costs and redistribution effects linked to PO1. The business-as-usual costs under the baseline scenario, related to the uncertainty to assess what liability rule would apply to AI and what burden-of-proof rule a court would apply in a concrete case, are higher than any potential adjustment costs borne by potentially liable parties. It will be easier for companies to estimate liability risks and related costs. While under the baseline scenario courts might apply on an ad-hoc basis alleviations of the burden of proof to remedy what they consider an unequitable result, clear and harmonised alleviations of the burden of proof will help liable parties to know what to expect in case AI is involved both domestically and cross-border. Companies operating cross-border would benefit from reduced compliance costs compared to the very fragmented baseline scenario. Such clarity might also help companies get appropriately priced liability insurance coverage.

In cases where the specific characteristics of AI would not have allowed the victim to prove the necessary facts under the baseline scenario, PO1 would shift the cost of compensating the relevant damage from the victim to the liable person, increasing the latter’s liability exposure. This result is not regarded as an undesirable impact or undue burden. It is in line with the policy objective to ensure that victims of damage caused with the involvement of AI systems have the same level of protection as victims of damage caused by other technologies and in general with the purpose of liability law. It also achieves a more efficient cost-allocation to the person best placed to prevent damage from occurring. For the impact analysis, it is taken into account as a re-distribution effect. These effects are inherent in the Commission’s objective of avoiding that victims are less protected due to the use of AI. In practice, the added burden on potentially liable parties is likely not to be substantial because:

-a major share is likely to have professional liability insurance 160 , which covers also risks linked to business activities involving AI 161 . It allows to limit costs to the annual insurance premium, making it even more predictable and easier to manage;

-PO1 does not involve a general reversal of the burden of proof, but only targeted adjustments to counter-balance the specific challenges of AI. The negative economic impacts for potentially liable parties are likely to be very marginal 162 ;

-PO1 is likely to achieve an overall more efficient allocation of the burden of proof, as the difficulty to establish how or why an AI system arrived at a harmful output is typically less burdensome for potentially liable parties having influenced the operation of that AI-system (e.g. developers, users) than for victims;

-as PO1 does not introduce new grounds of liability and keeps the basic allocation of the burden of proof intact, it is not expected to lead to a major increase in the number of civil actions – or an associated increase of insurance premiums – compared to the baseline, nor to significant familiarisation and implementation costs for businesses.

Especially the instrumental role of insurance in distributing and managing the impacts of the envisaged policy measures needs to be underlined. Ultimately, insurance coverage will allow potentially liable parties to spread liability costs across the community of all insured. This mechanism limits the economic burden on each individual insurance holder to the premium, preventing a possible deterring effect of liability risks and keeping market entry barriers low, which facilitates the roll-out of AI in particular by start-ups and other SMEs. 163 A large portion of potentially liable parties concerned will likely procure insurance coverage voluntarily to benefit from this cost-limiting effect. 164  

In this context, the regulatory framework established by the AI Act for the development and use of high-risk AI systems is likely to improve the conditions for AI risk assessment by insurers over the coming years. In addition, the proposed Data Act will promote access to data generated by a user’s product and thus facilitate the provision of services that depend on or can be improved by such data, including insurance and data analytics. 165 Moreover, the Commission services will respond to the EP’s call to work closely with the insurance market to develop innovative insurance products 166 and facilitate a dialogue between the insurance industry and companies active in the AI market (in particular SMEs).

Consequently, while a lack of actuarial data at both the current and the initial stage of implementation is likely to influence to some extent the premiums of AI-specific insurance products, for a transitional period, and make these premiums more volatile, this effect is expected to dissipate quickly as the data generated during the operation of these technologies will allow risk estimations to converge faster towards the optimum than in the case of ‘traditional’ technologies.

Hence, a development of a competitive insurance market for AI-related liability risks may safely be expected; and such development will provide the necessary conditions for effective coverage at moderate prices. 167 This expectation is supported by the fact that the insurance industry is forcefully pursuing the development of innovative AI-specific insurance products, to explore the substantial new opportunities linked to this growth market. 168 First insurance policies designed to cover specifically AI-enabled technologies have already been rolled out. 169 AI liability insurance may also be incorporated as an additional feature into existing general policies. 170  

With respect to high-risk AI-systems, PO1 could indirectly entail some minor administrative burden, namely to the extent that it prompts the disclosure, in particular in the framework of civil proceedings, of information documented pursuant to the AI Act. However, PO1 relates only to information that had to be logged, documented, and stored, in any case, for possible disclosure to supervisory authorities pursuant to the AI Act. This measure would thus firstly not affect private individuals or companies not falling in the scope of the AI Act. In addition, providing the same information, in particular also in the context of civil proceedings, is not expected to entail a significant added burden. Moreover, the harmonised disclosure rules would require national courts to ensure proportionality, i.e. to avoid any excessive administrative burden.

- Impacts on victims of damage caused by AI (natural persons and businesses): PO1 would relieve the victims of the burden of bearing the damage, to the extent that their claims for compensation would have failed under the baseline due to the specific challenges of AI. This burden would be re-distributed to the person responsible for causing the damage. This applies not only in respect of material damage, but also pure economic loss and non-material harm (such as psychological harm and damage caused by discrimination) to the extent that these types of harm are compensable under existing rules. PO1 would also reduce victims’ costs linked to the burden of proof (e.g. for expert analysis), by ensuring access to relevant information and ensuring that victims do not have to prove how or why an AI-system arrived at a certain output.

- Consumers: A faster roll-out of AI-technologies under PO1 would benefit consumers, e.g. in the form of faster and more personalised services, innovative and performant products as well as advances in the fields of health, safety, security, mobility, sustainability, circular economy, media, etc. Given the overall positive economic impacts also on businesses, it is not expected that the envisaged measures would lead to costs being passed on through increased consumer prices.

- Insurance companies: PO1 may marginally increase the take-up of insurance by potentially liable parties – provided the insurance coverage is not already included in existing all-risks-policies 177 . Increased legal certainty and reduced fragmentation create more favourable conditions for offering insurance coverage and awareness of liability risks may slightly rise due to this initiative. An increased coverage rate would benefit victims of damage as insurance claims provide an easier path to compensation and relieve victims of the liable party’s insolvency risk.

- Indirect economic impacts and impacts on the competitiveness of the internal market: Through the avoidance of liability gaps, PO1 would contribute to an efficient cost allocation. Its combined impacts are expected to have a positive effect on cross-border trade in AI-enabled products and services and the development of the European AI-sector as a whole. 178 The economic study estimated that a combination of alleviations of the burden of proof (as per PO1) and measures to harmonise strict liability for certain AI-enabled products and services (cf. PO2 and 3) would increase the cross-border trade in the AI-enabled goods and services falling under the six use-cases analysed in depth for that study by about 5 %. While PO1 does not include all of the assumptions made for that estimation, it is nevertheless relevant because the decisive drivers of the expected economic benefits – increased legal certainty, reduced fragmentation and increased consumer trust – are likely to materialise under PO1. 179

- Only small incremental impacts on enforcement, adjudication and litigation costs, borne by MS and parties to the proceedings respectively, are expected under PO1. The envisaged alleviations of the burden of proof are not likely to entail a substantial increase in the number of civil actions, as they are designed to apply only in the confined cases where the specific characteristics of an AI system make it unduly difficult to meet the default burden of proof. Moreover, the burden of proof will be distributed more efficiently overall, as potentially liable parties must by definition be capable of influencing the operation of AI-systems. They are therefore typically in a position to more easily discharge the burden of proof, with respect to how or why such systems arrived at a certain harmful output. This has a cost-cutting effect on overall enforcement, adjudication and litigation costs.

(c)     Coherence – in particular with the PLD review

As explained under 5.2.(a), PO1 would be coherent with the – not AI-specific – measures envisaged in the framework of the PLD review. These instruments are complementary as they address challenges posed by emerging digital technologies with respect to claims based on different grounds, directed against different liable persons and covering compensation for different victims and types of harm. The respective instruments use a consistent approach and similar tools (access to information, alleviations of the burden of proof) to ensure in their combination that products or services using AI or other digital technologies do not make it more difficult under any of the existing pillars of liability to get compensation compared to traditional products. The envisaged measures will together contribute to creating a more consistent and adequate liability framework, without upsetting the balance established by the existing rules.

PO1 would also be coherent with the proposals already adopted as part of the follow-up to the AI White Paper, in particular the AI Act. It would namely take over definitions of key concepts such as ‘AI-system’, ‘provider’ and ‘user’, although additional criteria would be necessary to ensure that the envisaged alleviations of the burden of proof apply only where the specific characteristics of AI effectively challenge existing liability rules. The provisions on disclosure of information and presumptions of causality would build specifically on the requirements of the AI Act. PO1 would thus complement this act providing an additional incentive for ensuring the safety of AI-enabled products and services as well as the respect of fundamental rights.

Finally, MS could fit the envisaged alleviations of the burden of proof into their national liability regimes without disrupting their respective legal traditions. This PO allows for sufficient flexibility and is based on tools that are already well-known in MS’ civil liability systems.

(d)     Proportionality

PO1 is limited to the measures strictly necessary to address the AI-specific problems identified. In particular, it would not touch upon the substantive conditions of liability like fault or causality, but focus on targeted proof-related measures ensuring that victims have the same level of protection as in cases not involving AI.

(e)     Considerations regarding the choice of instrument (binding v. non-binding)

A directive would be the most suitable instrument, as it could provide the desired harmonisation effect and legal certainty, while also being the adequate instrument to enable MS in a flexible manner to fit the harmonised measures without friction into their national liability regimes.

A mandatory instrument would prevent protection gaps stemming from partial or no implementation. It would thus likely ensure higher and more harmonised protection of victims compared to the baseline scenario. 187 While a non-binding instrument presents a less intrusive approach, it is unlikely to address the identified problems in an effective manner. The implementation rate of non-binding instruments is difficult to predict and there is no sufficient indication that the persuasive effect of a recommendation would be strong enough to produce consistent adaptations of national laws. This effect is even more unlikely for measures concerning private law, of which extra-contractual liability rules form part. This area is characterised by long-standing legal traditions, which makes MS reluctant to pursue harmonised reform unless driven by the clear prospect of internal market benefits under a binding EU instrument 188  or the need to adapt to new technologies in the digital economy. Moreover, the significant existing divergences between MS’ liability frameworks (see 2.4.), are another reason why a recommendation is unlikely to be implemented in a consistent manner.

Divergences and reduced legal certainty in cross-border cases would thus persist even to the – likely very limited – extent that MS choose to implement a recommendation. A selective and inconsistent implementation would solve neither the problem of legal uncertainty nor that of impending fragmentation and it would not ensure consistent victim protection throughout the EU. The economic study confirmed that a non-binding initiative would not address the identified internal market obstacles effectively, as the underpinning problems will likely be perpetuated to a substantial extent. 189  It concluded that a non-binding instrument would not achieve any increase in cross-border trade. 190 Due to this likely lack of effectiveness, choosing a non-binding instrument would not be in line with the principle of proportionality, which requires the choice of the least burdensome PO, which is still suitable to achieve the policy purpose.

While the above considerations clearly militate in favour of implementing PO1 by way of a binding legal instrument, that instrument has to allow Member States to integrate the harmonised measure into their traditional legal systems without creating frictions. A directive would therefore be the most appropriate tool to achieve the policy objectives while maintaining the necessary flexibility to preserve the various established national approaches and legal traditions in the politically sensitive field of civil law.

6.2. PO2: PO1 + strict liability for AI use-cases with a specific risk-profile, possibly coupled with mandatory insurance

PO2 differs from PO1 as regards the strict liability regime applicable to users of AI technologies with a special risk-profile, possibly coupled with a mandatory insurance regime. The following impact analysis therefore focuses on these elements.

(a)Effectiveness

- Degree to which SO1 and 2 191 would be achieved: A harmonised strict liability regime, possibly coupled with mandatory insurance, is a suitable instrument to ensure legal certainty and prevent fragmentation, as MS would have to implement a consistent minimum strict liability standard with respect to the covered activities. Provided that the AI-enabled technologies covered by that regime can be specified with precision, the companies operating/using those technologies could have an even clearer and consistent basis for assessing their liability risk. However, the time horizon of the roll-out of technologies with a relevant risk profile and degree of autonomy as well as their features which allow to define them in a legal act with sufficient certainty are not yet known. It would therefore, at this point in time, be to a certain degree difficult to define the risk profile of those technologies with the desirable degree of precision, and to specify them in a legislative instrument in a way that ensures the legal certainty necessary for the definition of the scope of the harmonised strict liability regime. This consideration would apply also to a possible mandatory insurance regime covering strict liability. The obligation to ensure insurance coverage is a market entry requirement. This means that it is crucial to enable market participants to assess with a high degree of certainty whether they fall under this requirement or not.

Moreover, because of the uncertainty related to the risk profiles explained above, it cannot be easily assessed at this stage if having an alleviation of the burden of proof (PO1) would not be sufficient to address the identified problems. This means that the effectiveness of the strict liability element of PO2 is also reduced.

For these reasons, an instrument including strict liability for the use/operation of certain types of AI-enabled technologies would likely be less effective than it should be, at this point in time, in achieving specific objectives 1 and 2 than PO1.

- Degree to which SO3 would be achieved: A harmonised strict liability and possible mandatory insurance regime, as the distinguishing features of PO2, could prevent a lack of compensation even more effectively than the alleviations of the burden of proof common to PO1 and 2. The expected effects on societal trust follow a similar pattern as under PO1, as strict liability represents simply another – potentially even more effective – way of ensuring an effective compensation of victims. As with the alleviations of the burden of proof, this is likely to have a positive effect on consumers’ perception of the appropriateness of liability rules, as well as their trust in AI applications. 192 This in turn is likely to increase their willingness to take up AI-enabled products and services. 193 At the same time, it would need to be acknowledged that pending the market maturity of relevant AI-enabled products and services, the specific modalities of their operation and the degree of control exercised respectively by the relevant stakeholders over the risks linked to the operation are at this point in time not known yet.

- Indirect social impacts: The ability of strict liability rules to incentivise safety efforts depends to a large extent on whether the strictly liable person has cost-efficient means to prevent damages. 194 The control criterion envisaged for assigning strict liability to professional users/operators is therefore conducive to the desired incentive effects.

Moreover, behavioural research has shown that a strict liability framework for AI is more likely to be perceived as predictable and transparent than fault-based liability. 195 The strict liability element is thus likely to increase societal trust in the justice system.

(b)     Efficiency

- Impacts on businesses (in particular as potentially liable parties): Provided that the AI-enabled technologies covered by the harmonised strict liability can be determined with a sufficient degree of precision, this element of PO2 would be suitable to increase legal certainty and reduce fragmentation by reducing legal information/representation, internal risk management, and other compliance-related costs, and generate additional cross-border revenue. 196 Moreover, the expected positive impacts of PO2 on consumers’ willingness to take up AI-enabled products and services (see below) would directly or indirectly benefit all companies in the AI value chain. Start-ups and other SMEs would benefit to a higher degree from the overall positive economic impacts of PO2. 197

In certain cases, beyond merely harmonising existing strict liability rules, PO2 would entail the application of strict liability and possibly an insurance obligation to users of AI technologies that would otherwise be subject only to fault-based liability under national law. In such cases, PO2 could be a dis-incentivising factor for businesses choosing between AI-enabled technologies and functionally equivalent alternatives. However, this effect is likely offset by the cost reduction and internal market opportunities generated through harmonised liability rules. The economic study found that the moderate compliance costs linked to PO2 would be outweighed by cost savings thanks to higher legal certainty, saved resources on compliance, and higher revenue enabled by a clearer and less fragmented legal framework. 198  

Moreover, the role of insurance solutions, as described under PO1, is instrumental in this respect, as it limits potentially liable parties’ costs to the insurance premiums, keeping market entry barriers low. 199   In principle, this mechanism holds true whether the relevant risk is covered by voluntary (market-driven) insurance or a mandatory insurance regime (either harmonised or regulated at national level). 200 In the public consultation, insurance stakeholders have pointed to a lack of statistical data on accidents and damages, which could initially drive risk margins and thus insurance premiums. Some SME stakeholders have also raised concerns about high insurance premiums due to the difficulties in assessing the covered liability risks. On the other hand, the fact the insurance industry is already proactively developing new insurance products for AI risks 201 and that this field is considered as the next big growth market 202  confirm the expectation that coverage will be offered at competitive prices. An additional important factor is the parallel initiative adapting the PLD to the digital age. Based on a (statutory or contractual) subrogation of the victim’s claim to the insurer, it will facilitate it for insurers to take recourse against producers, where defective AI software caused or contributed to the insured damage. This would allow insurers having compensated the victim’s damage to claim the compensation back from the insurer, thus also contributing to keeping premiums low. 203 In any event, the Commission services will respond to the Parliament’s call to work closely with the insurance market. 204 The Commission will notably facilitate a dialogue between the insurance industry and companies active in the AI market (in particular SMEs).

- Victims of damage caused by AI applications with a special risk profile (natural persons and businesses): A harmonised strict liability regime would facilitate victims’ access to compensation to an even greater extent than alleviations of the burden of proof. A possible mandatory insurance regime would relieve victims’ of the liable parties’ insolvency risk and provide them with an even cheaper, faster and easier path to compensation. 206  

- Consumers: The expected benefits in terms of a faster roll-out of AI-enabled products and services are likely to be similar under PO2 as under PO1, because their common element – alleviations of the burden of proof under fault-based liability rules – applies to the major part of AI-enabled products and services affected by the initiative. The strict liability and mandatory insurance elements of PO2 are much more limited in scope. In respect of these elements, the difficulty linked to specifying already at the present point in time the AI-enabled technologies with a ‘strict liability profile’ could lead to a certain degree of legal uncertainty, which could potentially dis-incentivise AI rollout in certain cases.

- Insurance companies: The harmonisation of strict liability under PO2 is likely to entail some additional demand for insurance, in particular when coupled with a compulsory insurance regime, which generates new market opportunities for insurance companies.

- Indirect economic impacts and impacts on the competitiveness of the internal market: While the envisaged strict liability regime may be the most certain and easiest way to ensure that the victim does not bear the cost of the damage, it may not in all cases lead to the cost allocation on the party which was at the origin of the damage. However if it is coupled with mandatory insurance, such a cost-allocation will be achieved through subrogated recourse claims. 208  In case the insurance company of the strictly liable person compensates the victim, a damage claim of the victim based on fault or product liability would be re-assigned to the insurance company. On the basis of this claim, the insurance company could have recourse against the person at the origin of the damage, for instance against the producer if the product was defective. Thereby the most efficient cost allocation would be achieved. This scenario could be extended along the value chain. For example, if the producer has a liability claim against the developer, the insurance company of the producer would have a recourse possibility on the basis of this contractually subrogated claim against the developer. In the end, the insurance of the developer would pay. At the same time, insurance coverage of liable persons would limit the economic costs to the annual insurance premium and keep the market entries for AI producers and operators low, while it would be ensured at the same time that the victim’s harm would be compensated in a smooth way.

- For similar reasons as under PO1, only small incremental impacts on enforcement, adjudication and litigation costs are expected under PO2.

To the extent that risks within the scope of the strict liability element of PO2 are not already covered by a national strict liability regime under the baseline, the easier and more predictable path to compensation afforded by this measure may lead to a slight increase in the number of claims made. Any such possible increase is likely to be marginal. As strict liability would apply only in cases involving significant risks to important legal interests (life, health, property), victims would in such cases be likely to seek compensation also under the baseline, despite the challenges and costs linked to fault-based claims. 210

(c)    Coherence – in particular with the PLD review

PO2 would complement the measures envisaged in the PLD review. In particular, the liable persons defined by law are different and their liability is based on a different relationship they have to the risk. The law would attribute liability to the operator/user under the harmonised strict liability regime because the operation of the AI system would realise the risk for the victim. At the same time, the operator draws a benefit from the operation. The law also attributes liability to the producer under the (revised) PLD, despite the fact that the producer is a step more removed from the contact with the victim. The producer is made liable because of producing a defective product (including the AI that is necessary for it to work).

In cases where the producer is nevertheless at the same time also the user/operator, the fact that they would be subject to the revised PLD and the harmonised strict liability regime for AI would not lead to inconsistent results because these liabilities would intervene in different capacities. Such scenarios already exist in the national laws where national strict liability schemes and product liability apply in parallel.

As regards national law, the coherence of PO2 with the existing legal framework could be questioned in cases where the use of AI-enabled technologies would be subject to the harmonised strict liability regime while the use of functionally equivalent non-AI technologies is not subject to strict liability under national law.  (d)     Proportionality

At this point in time, it is not clear whether, despite the alleviations of the burden of proof envisaged under PO1, , there is a liability gap which would need to be filled with a harmonised strict liability regime for AI operators. It is possible that the less interfering measures under PO1 will be sufficient to achieve the policy objectives. Moreover, at this point in time there is still some uncertainty regarding the precise characteristics, risk-profile and conditions of deployment of AI-enabled technologies with a potential ‘strict liability profile’. Therefore the proportionality of PO2 cannot be affirmed with certainty at this point in time.

(e) Considerations regarding the choice of instrument (binding v. non-binding)

The considerations made in the context of PO1 apply to an even larger extent: As a harmonised strict liability and mandatory insurance regime involves more far-reaching adaptations of MS’ liability rules, a non-binding instrument is not suitable to ensure a consistent implementation of PO2. For the same reasons as set out under PO1, a Directive seems to be the most appropriate legislative tool.

6.3. PO3: PO1 + targeted review regarding the strict liability and mandatory insurance elements of PO2 (staged approach) 

(a) Effectiveness

During the first stage, PO3 would be equally effective to PO1. The second stage, i.e. the targeted review mechanism, would systematically re-assess the need for the strict liability and mandatory insurance elements of PO2. PO3 would ensure that the necessary information for this re-assessment is available. The review would assess if the technological and market developments as well as empirical evidence on AI liability cases confirm the need for such measures. This mechanism does not pre-empt the necessary future policy decision on the harmonisation of strict liability and mandatory insurance. But if there were a policy decision to that effect, the targeted review mechanism would have prepared the groundwork for realising the potentially even greater effectiveness described under PO2. The second stage would also gain more certainty about the technological and regulatory context defining the risk-profile and conditions of deployment of the AI-enabled technologies with a potential strict liability profile. This would thus allow to specify the material and personal scope of a possible strict liability regime with greater precision. It would therefore be conducive to ensuring legal certainty and uniform implementation (SO1 and 2).

(b) Efficiency

During the 1st stage, PO3 would have the same impacts as PO1. As MS anyway have to report on the implementation of the initiative in line with better regulation requirements on evaluation and monitoring, targeted reporting requirements supporting the review mechanism under PO3 would not entail a significant burden for them.

The potential benefits of a harmonised strict liability regime would materialise later than under PO2. If there is a liability gap which would have needed to be filled with a harmonised strict liability regime for AI operators, the disadvantages of such liability gap would have their effect during a length of time corresponding to the period of data collection and assessment of the need for strict liability. However, the following factors will improve the efficiency of the measures potentially taken at the second stage:

-By ensuring that the assessment of a possible strict liability regime for the use of AI-enabled technologies can rely on a more developed factual basis, the staged approach minimises the risk of creating an uneven playing field for AI-enabled technologies.

-The staged approach allows the development of tailored market-driven insurance solutions, which can be taken into account at the second stage when assessing the need for and effects of a mandatory insurance regime. Moreover, AI-technologies potentially subject to such a regime will be rolled out over the coming years and statistical accident data will be accumulated. Therefore the lack of actuarial data as main potential cost-driver of AI-specific liability insurance will have become considerably less relevant by the time of the targeted review under PO3. 

-As relevant safety standards for the AI-enabled technologies will be available by the time of the targeted review, the conditions for assessing liability risks will have improved for both insurers and liable users/operators.

(c)    Coherence

PO3 would insert itself without friction into the existing liability system and be consistent with the other strands of the Commission’s AI policy. The staged approach would allow the Commission to take stock of the practical effect of the planned adaptations to the PLD, in particular as regards providers of safety-relevant AI-systems. The targeted review mechanism would allow the Commission also to ensure coherence with future AI-related policy measures beyond the proposed AI Act. This may for instance concern future safety rules tailored to AI-systems with a ‘strict liability profile’ like autonomous vehicles and highly autonomous AI-enabled drones.

(d)     Proportionality

PO3 would ensure that future technological, regulatory and jurisprudential developments will be systematically taken into account to verify the need to harmonise strict liability for certain uses of AI. The staged approached is thus strictly based on the principle of proportionality.

(e)     Considerations regarding the choice of instrument (binding v. non-binding)

6.4.For the same reasons as set out under PO1 and PO2, a binding legal instrument – a Directive – would be the most appropriate tool to implement PO3. Evolution of policy options

It is important to note that the policy options retained for detailed assessment have evolved due to the conclusions of the economic study, as well as due to the results of the public consultation and following discussions with stakeholders.

The economic study assumed two different combinations of alleviations of the burden of proof under fault-based liability rules with a harmonised strict liability regime    and, by consequence, did not explicitly assess the economic impacts of alleviations of the burden of proof, as per PO1, taken in isolation.

However, it proved important to consider alleviations of the burden of proof on their own (PO1), as an alternative to introducing these alleviations together with strict liability (PO2). We consider that the assessments made by the economic study are still largely relevant because:

- According to the study, the economic benefits of an EU initiative on AI liability are primarily attached to the expected gains of legal certainty and reduced legal fragmentation. These effects are expected to materialise also with respect to the alleviations of the burden of proof taken in isolation (PO1 and first stage of PO3), which will clarify in a harmonised manner how the burden of proof is to be handled in cases involving AI.

- The measures that PO1 shares with two of the policy options assumed for the purposes of the economic study are relevant for the major share of AI-enabled products and services, and thus decisive for the economic impacts on most stakeholders. This is because only a small set of AI-enabled technologies would have a risk profile warranting the application of strict liability.

6.5.UN Sustainable Development Goals and indirect environmental impacts

The policy options are expected to contribute to the rollout of AI and thus to achieving the related SDGs and targets 211 . They would also impact positively on them by contributing to the enforcement of the AI Act, because effective legislation on transparency, accountability and fundamental rights will direct AI’s potential among other benefits for individuals and society towards achieving the SDGs.

As regards indirect environmental impacts, all policy options are expected to contribute – albeit to a non-quantifiable extent – to the uptake of AI applications that are beneficial for the environment. For instance, AI systems used in process optimisation make processes less wasteful (e.g. by reducing the amount of fertilizers and pesticides needed, decreasing the water consumption at equal output, etc.). However, there is no sufficient basis for assessing the environmental impact, since this basis could only relate to actually marketed AI-enabled products/services. However, this initiative will improve, to a great extent, the market roll-out conditions of AI-enabled products/services, which are not yet on the market.

In addition, AI systems supporting improved vehicle automation and traffic management contribute to the shift towards cooperative, connected and automated mobility, which in turn can support more efficient and multi-modal transport, lowering energy use and related emissions. On the other hand, possible unintended effects cannot be excluded. For instance, an increase in traffic may be possible that could partly offset the lower energy use and emissions achieved through more efficient and multi-modal transport.

7.How do the options compare?

The PO were compared by way of a multi-criteria analysis (MCA) taking into account their effectiveness, efficiency, coherence and proportionality. 212  The following impact matrix compares the scores of the options for all of the main impact assessment criteria, using simple aggregation and assuming an equal weight of each individual criterion. The respective scores flow from the expected impacts as presented under heading 6 above. This ranking is linked primarily to the fact that PO2 includes a strict liability element which, for the reasons explained in section 6.2. a), lowers the effectiveness of this option at this point in time. Therefore,  PO2 scores slightly lower in terms of its effectiveness to achieve specific objective 1 (increase legal certainty regarding liability for AI), coherence and proportionality, leading to a marginally lower overall score.

As three out of these six individual impact assessment criteria would come under the umbrella of effectiveness, effectiveness is de facto given more importance than the other IA criteria. Therefore, sensitivity analysis was carried out.  The results of the MCA and sensitivity analysis are summarised in the following table 213  which shows that PO1 and PO3 rank highest. This result is consistent with the fact that they involve the same harmonising provisions at the present stage. The feature distinguishing these options, the targeted review mechanism, does not lead to different impacts of the implemented measures to be implemented at this first stage.

8.Preferred option

8.1. Selection of the preferred option: PO3 (staged approach)

While liability rules are instrumental in controlling the risk associated with the development and use of AI, the findings of the research and the results of the public consultation informing this impact assessment have shown that the specific features of the AI will challenge their application. They also point towards a staged approach, as the best way to serve the objectives of this initiative, accommodate the concerns of the public and the industry, and at the same time allow Member States to integrate the adaptations into their existing traditional systems without frictions.

Specifically, a clear majority of business stakeholders opposed strict liability in the public consultation. In addition, the relevant insurance market is still under development due to the currently lower predictability of AI-related risks linked to the lack of sufficient data.

In this respect, the multi-criteria analysis provides a clear ranking between PO1/PO3 and PO2, confirming that it is preferable not to lay down a harmonised strict liability regime for certain types of AI-enabled products and services at the present time, whether or not coupled with mandatory insurance.

In addition and in order to inform the necessary political decision between PO1 and PO3, it is also important to consider how well these options take into account the suggestions of the European Parliament as well as stakeholder opinions. In this respect, firstly, only the staged approach (PO3) incorporates both main elements suggested by the Parliament – a facilitated burden of proof under fault-based liability rules and a limited strict liability regime for certain AI-enabled technologies – explicitly into the proposed legislative instrument. While the strict liability element is not yet implemented at the present stage, the targeted review mechanism provides a dedicated framework preparing the ground for the future policy decision on this element. Secondly, by explicitly acknowledging the possible need for a harmonised strict liability regime in the proposed legislative provisions, PO3 also satisfies to a greater extent the opinions expressed by non-business stakeholders, a large majority of whom supported the harmonisation of strict liability.

If, at the stage of the review, a political decision is taken to propose a harmonised strict liability regime, this measure could be designed in a way to meet concerns expressed by business stakeholders, which are at present sceptical towards it.

In light of these considerations, the staged approach (PO3) is the most balanced, politically feasible, proportionate and yet effective option. It is most suitable to deliver the desired economic benefits in terms of roll-out of AI-enabled products and services in the internal market, and to increase citizens’ trust in AI by ensuring that victims who suffered harm caused with the involvement of AI systems enjoy the same level of protection as victims who suffered harm caused by other technologies. It is also most adapted to the political context of the AI liability initiative, including the Parliament’s legislative own-initiative resolution, and stakeholder feedback.

8.2. Rationale and main impacts of the preferred PO 

PO3 would achieve the SO of reducing legal uncertainty, preventing fragmentation and protecting victims by ensuring an effective path to compensation. The measures included in PO3 would also incentivise compliance with safety requirements and requirements designed to safeguard fundamental rights, applicable to AI-enabled products and services and thus reinforce the eco-system of trust in AI.

The main impacts concern potentially liable persons and potential victims. On the one hand, the internal market obstacles identified will be reduced and this will lead to reduced costs for business, which would outweigh any possible adaptation (substantive compliance) costs. On the other hand, PO3 will reduce victim’s costs and re-distribute the burden of bearing the damage to the person responsible for causing it.

The choice of PO3 as preferred option is a result of the application of the proportionality principle. This PO will allow to reassess the need for harmonising, in addition to aspects related to the burden of proof, also situations with specific risk profile where the use of AI would warrant strict liability. At that point in time, the price of insurance will likely be more stable because of more available data.

8.3. The interplay with the PLD review: consistency and complementarity of the package of measures for the compensation of damage caused by AI

As explained above, the measures from the preferred PO of the two impact assessments address challenges posed by emerging digital technologies with respect to claims based on different grounds, covering different types of damages and victims and against different liable persons. Both preferred PO have a consistent approach and use similar tools (access to information, adaptations of the burden of proof) to ensure, in their combination, that victims of damage caused by AI-systems have the same level of protection compared to damage caused by traditional products, no matter the liability path taken.

8.4. One-In-One-Out

The preferred PO involves small incremental adjustment costs for liable parties. These costs are linked to a transfer of costs of compensation from victims. The envisaged alleviations of the burden of proof will allow victims to claim compensation successfully in cases where they would have been unable to prove a justified claim under the baseline scenario due to the AI-specific difficulties of proof. This transfer is fully in line with the policy objectives and the spirit and purpose of liability rules. Its effect on liable parties has been approximated in terms of possible increases of general liability insurance premiums for a transitional period, where data is lacking. Also, potentially liable entities are in many cases likely to be covered by voluntary insurance, so that their exposure –which will be in any case done at their own interest- would again be limited to the premiums. Overall, it is expected that the annual volume of premiums paid for liability insurance will increase by EUR 5.35mln - 16.1mln due to the preferred option. This incremental increase does not represent a significant burden for stakeholders active in the AI market, given that the overall amount of these premiums is over EUR 42bln per year, since it only represents an increase of 0,013% - 0,04%

This cost transfer will be mostly relevant for businesses as potentially liable parties rather than for natural persons. This is because the AI-specific liability gaps addressed by the preferred policy option are more likely to affect the liability exposure of actors with an active influence on the functioning of the relevant AI systems. The benefits of the policy options for companies in terms of increased legal certainty and reduced legal fragmentation outweigh these marginal adjustment costs, because the business-as-usual-costs under the baseline scenario related to the uncertainty about what liability would apply to AI and what burden-of-proof rule a court would apply in a concrete case are higher.

The preferred PO will not introduce any administrative requirements (e.g. reporting, registration, monitoring) for any of the entities within its scope, i.e. potentially liable parties or victims. A possible incremental increase in judicial claims and the related costs borne by the parties or the courts are not included in the ‘administrative costs’ for the purposes of the One-In-One-Out approach, as they are not directly linked to the compliance with a law.

The preferred policy option will not entail costs incurred in related markets or experienced by stakeholders that are not directly targeted by the initiative. In particular, as the initiative is expected to generate net cost savings for businesses active in AI, it is not expected to lead to increased consumer prices.

9.How will actual impacts be monitored and evaluated?

In order to ensure that the targeted review mechanism under the preferred PO (staged approach) can rely on a sufficient evidentiary basis, this mechanism could:

-provide for reporting and information sharing by MS regarding the application of the measures under PO 1 in national judicial or out-of-court settlement procedures;

-use information collected by the Commission or market surveillance authorities under the AI Act (in particular Article 62) or other relevant instruments;

-use information and analyses supporting the evaluation of the AI Act and the reports to be prepared by the Commission on the implementation of that act;

-take into account any information and analyses supporting the assessment of relevant future policy measures under the ‘old approach’ safety legislation 214 ;

-rely on the information and analyses supporting the Commission’s report on the application of the Motor Insurance Directive with regard to technological developments (in particular autonomous and semi-autonomous vehicles) pursuant to its Article 28c(2)(a).

To evaluate the effectiveness of the preferred PO, success criteria and data sources have been defined on a provisional basis for each of the SO. 215  For instance:

-for SO1, effectiveness will be evaluated based on the level of legal certainty as perceived by business stakeholders, compared to the results of the public consultation;

-for SO2, the adoption of any diverging measures on AI liability, as reported by MS, will be taken into account to evaluate the initiative’s effectiveness;

-for SO3, the degree to which compensation gaps have been prevented and consumer trust has increased will be assessed based on expert and behavioural analysis, and MS reporting.

Efficiency would be evaluated against specific operational objectives, for example:

-With respect to companies, cost reductions due to increased legal certainty and reduced fragmentation could be used as benchmark, based on stakeholder feedback.

-Regarding victims, the level of difficulty encountered when claiming compensation for damage caused by AI could be compared to the experts’ estimates in the Economic Study.

-For insurers, the liability-related conditions for offering insurance coverage could be assessed, based on feedback from a stakeholder survey and/or a dedicated workshop.

In addition, criteria for evaluating the proportionality, coherence and continued relevance as well as the EU added value of the policy measures have been developed. For instance, the emergence of any more effective or efficient means to achieve the policy objectives would be assessed by Commission services taking into account stakeholder feedback and information reported by MS on relevant legal cases. Moreover, the synergetic interplay between this initiative and the AI Act as well as the revised PLD would be assessed.

Annex 1: Procedural information

1.Lead DG, Decide Planning/CWP references

DG Justice and Consumers, PLAN/2020/9848 - Adapting liability rules to the digital age and Artificial Intelligence, CWP 2020

2.Organisation and timing

Roadmap consultation period – 30 June to 28 July 2021

Open public consultation period – 18 October 2021- 10 January 2022

ISSG meeting on the Impact Assessment – 18 February- participants: SG, LS, JUST, GROW, CNECT, JRC, SANTE, EMPL, COMP, FISMA, AGRI, ECFIN, ENV

3.glossary

4.Consultation of the RSB

4.1. Upstream meeting with the RSB – 30 March 2020

The guidance and advice provided by the RSB was implemented in this impact assessment, in particular:

-The interplay of this planned initiative with other relevant policy measures (in particular the PLD review and the AI Act) were addressed in detail in close cooperation with other Commission services, to ensure that related initiatives fit into a coherent overall intervention logic, serve consistent objectives and achieve synergies. Explanations on these aspects are provided throughout the impact assessment as well as in Annexes 6, 7 and 8.

-The measures suggested in the European Parliament’s resolution on a civil liability regime for AI were to a large extent incorporated into the assessed policy options. For detailed explanations regarding the degree of alignment with the Parliament’s resolution, see Annex 4, section 4.

-The lack of quantified data (mainly due to the fact that the relevant AI-enabled products and services are for the most part not yet on the market) is acknowledged. The measures undertaken to remedy the scarcity of data and to nevertheless generate quantified estimates are described. See in particular Annexes 4 and 10.

-The specific challenges of AI regarding existing civil liability rules are explained in detail in the impact assessment and Annex 5, and illustrated with examples. Concrete use-case scenarios demonstrating these challenges have been elaborated together with AI experts from the Joint Research Centre (Annex 13). In this respect, the communalities and differences between the Product Liability Directive and national civil liability rules are explained.

-The role of insurance solutions has been emphasised and addressed in detail in the assessment of policy options, taking into account detailed discussions with insurance experts and stakeholders, as well as the feedback received from these stakeholders during the public consultation and other consultation activities.

-The preferred option was conceived specifically to enable an effective monitoring and evaluation of the envisaged measures, and to ensure that policy decisions on the more far-reaching measures considered (strict liability, possibly coupled with mandatory insurance) will be taken, at the stage of the targeted review, on an even more comprehensive evidence base.

4.2. Opinion of the RSB and responses

The Impact Assessment report was reviewed by the Regulatory Scrutiny Board. It received a positive opinion with reservations. The Impact Assessment has been revised to take into account the Board’s comments:

5.Evidence, sources and quality

The discussion on liability for AI has started with the work undertaken by the Expert Group on Liability and New Technologies 216 , which had a series of meetings in 2018 and 2019 and was organised in two formations: Product Liability and New Technologies. The New Technologies formation of the Expert Group adopted a Report in 2019 on the topic of liability for artificial intelligence and other emerging digital technologies. 217 This report informed the Commission services in the preparation of the Commission Report on AI Liability 218 accompanying the White Paper on AI.

In view of preparing this Impact Assessment, the Commission also contracted legal experts for further comparative legal analysis 219 , an additional study to provide economic analysis 220 and a third study focusing on behavioural analysis. 221

The specific details of all these studies, their scope and methodology are described in Annex 4.

In addition, the Commissions services also considered the ‘Comparative study on national rules concerning non-contractual liability, including with regard to AI’ which accompanied the European Added Value Assessment on a Civil liability regime for AI 222 commissioned by the European Parliament.

The Impact Assessment was further based on the results of the public consultation, the input provided by stakeholders on the roadmap, online webinars, survey with citizens (over 8,000) done in the context of the behavioural economics study and numerous bilateral contacts with stakeholders and meetings with public authorities. Annex 2 provides more details about these sources.

Annex 2: Stakeholder consultation

A. Outline of the consultation strategy/process 

In line with the Commission’s Better Regulation requirements, an extensive consultation strategy has been implemented to ensure a wide participation throughout the policy cycle of this initiative. The consultation strategy was based on both public and targeted consultations. The Commission has sought a wide and balanced range of views on this issue by giving the opportunity to all relevant stakeholders to express their views. In particular, the stakeholders addressed by the strategy were:

-EU and national consumer associations and civil society organisations active in the justice field, victims associations and individual victims of damage caused by products

-Industry associations at EU, national or sectoral level

-Economic operators, including SMEs (e.g. manufacturers of products and components, software developers, AI programmers, medical devices manufacturers, manufacturers of pharmaceutical products, retailers (offline and online), online marketplaces, repair and refurbishment providers, providers of services, operators and users of various products)

-Insurance associations and individual insurers

-Legal firms and lawyers

-Academic experts and research bodies

-Citizens

-National authorities.

The extensive consultation process included various activities, for which the content and results are described in greater detail in the subsequent sections of this Annex, along with explanations about how the received feedback was taken into account for the development of policy options covered by this IA:

-public consultation on the White Paper on AI and the Commission report on safety and liability;

-online interactive webinars with a broad range of stakeholders, focused on specific questions regarding AI liability;

-gathering of feedback on the published inception impact assessment;

-surveys and targeted interviews carried out in the framework of two supporting studies (general economic and behavioural analysis);

-a dedicated 12-weeks public consultation on Adapting Civil Liability Rules to the Digital Age and AI, which covered both the general review of the Product Liability Directive and the AI-specific issues addressed by this IA;

-a workshop with Member State representatives organised to gather feedback on the AI liability initiative and the review of the PLD.

The results of the dedicated public consultation are summarised first (Section B), and other consultation activities subsequently (Section C).

B.    Public consultation on Adapting Civil Liability Rules to the Digital Age and AI

1.    Introduction

The public consultation on Adapting Civil Liability Rules to the Digital Age and AI covered both the general review of the Product Liability Directive and the AI-specific issues addressed in this IA. Section I of the public consultation aimed to gather stakeholders’ views on how to improve the Product Liability Directive (PLD); this section received a total of 291 responses. Section II concerned problems linked to certain types of AI that make it possibly difficult to identify the potentially liable person, to prove that person’s fault and/or the defect of a product and the causal link with the damage. There was a filter question asking respondents to choose if they want to continue with Section II or finish the survey after Section I, so only 233 responses were provided in Section II. A large number of business associations (63), representing a wide variety of interests and company sizes, contributed as well as individual companies (29), including SMEs (9). Associations representing consumers also contributed (7), including BEUC. The consultation includes also input from individual citizens (95) and NGOs (10), as well as research institutions (14) and national ministries (5) 223 . Some questions received less answers, since not all participants answered all the questions. The following summary focuses on the results relevant concerning AI-related questions in particular (section II of the public consultation survey).

In terms of geographical representation, the consultation includes contributions from 21 Member States, as well as from third countries. The geographical coverage, however, is broader because some associations, which indicated Belgium as their country of origin, also represent stakeholders from Member States not directly mentioned in the responses. The majority of contributions come from Germany (91), followed by Belgium (39) and France (20). No campaigns were identified.

There were also 70 position papers submitted separately in the public consultation.2.    Short summary of key results

2.1.    Problems and problem drivers

(a) The specific characteristics of AI make it difficult to meet the burden of proof

The results of the public consultation confirm the existence of this problem driver:

Amongst responding consumer organisations, NGOs, national ministries, academic/research associations and EU citizens, an overwhelming majority agreed that:

(I)it could be difficult to link damage caused by highly autonomous AI to the actions and or omissions of a human actor (91%, or 85 out of 93, agreement amongst responding citizens; unanimous agreement amongst responding consumer organisations, NGOs, national ministries and academic/research associations),

(II) in the case of opaque and complex AI, it could be difficult for victims to prove that the conditions of liability are fulfilled (84%, or 78 out of 93, agreement amongst responding citizens; unanimous agreement amongst responding consumer organisations, national ministries, NGOs and research institutions), and

(III) because of AI’s specific characteristics, victims may in certain cases be less protected that victims of damage that did not involve AI (79%, or 74 out of 94, agreement amongst responding EU citizens; 84% agreement amongst responding consumer organisations, national ministries, NGOs and research institutions, with only one disagreeing).

Responding business stakeholders (business associations and companies/business organisations) also agreed broadly with this problem driver, but were less unequivocal:

-An absolute majority agreed that because of AI’s specific characteristics, victims may in certain cases be less protected that victims of damage that did not involve AI (54 %, or 45 out of 85, v. 26 %, or 22 out of 85, who disagreed).

-A relative majority agreed that in the case of opaque and complex AI, it could be difficult for victims to prove that the conditions of liability are fulfilled (41 or 35 out of 85 % v. 28 % or 24 out of 85 who disagreed).

-Only with respect to the statement that it could be difficult to link damage caused by highly autonomous AI to the actions and or omissions of a human actor , more business stakeholders disagreed (45% or 38 out of 85) than agreed (36% or 30 out of 85).

(b) Lack of compensation leads to a lack of trust in AI

A clear majority of overall respondents confirmed that the lack of adaptation of existing liability rules to AI may negatively affect trust in AI (60 %, or 135 out of 227, agreement v. 25%, or 58 out of 227, who disagreed) and the uptake of AI-enabled products and services (56%, or 126 out of 227, agreement v. 27%, or 60 out of 227, disagreement). In particular, responding NGOs, consumer organisations, academic/research institutions and EU citizens overwhelmingly confirmed these problems (92%). By contrast, close to 60 % (or 49 out of 85) of responding business stakeholders (business associations and companies/business organisations) did not confirm them, while around 27 % (or 23 out of 85) did. 

(c) Legal uncertainty

The public consultation confirmed legal uncertainty leading to internal market obstacles: Amongst responding consumer organisations, NGOs, academic/research associations and EU citizens, an overwhelming majority agreed that:

(I)it is uncertain whether and how liability rules under national law apply to damage caused by AI (84%, or 27 out of 32, agreement amongst consumer organisations, NGOs and academic/research associations; 85%, or 80 out of 94, amongst EU citizens),

(II)it is uncertain how national courts will address possible difficulties of proof and liability gaps in relation to AI (81%, or 25 out of 31, agreement amongst consumer organisations, NGOs and academic/research associations; 91%, or 84 out of 92, amongst EU citizens).

Responding business stakeholders were more divided: A bigger share disagreed with statement (i) (52%, or 44 out of 85), while still a sizeable share of 32 % (or 27 out of 85) confirmed uncertainty as to how national liability rules will apply. More business respondents agreed with the uncertainty as to how national courts will react to the challenges of AI: 34% (29 out of 85) confirmed uncertainty, while 43% (36 out of 85) did not.

Substantial shares of businesses stakeholders (although not the overall majority) think that the lack of adaptation of the current liability rules to AI will entail a number of negative economic consequences. They expect for instance that this lack of adaptation of liability rules will entail additional costs (e.g. legal information costs, insurance costs) for companies (27% or 23 out of 85), and that companies may defer or abandon certain investments in AI technologies if the current liability framework is not adapted (20% or 17 out of 85).

A slight majority of business associations representing primarily SMEs think that, if the liability framework is not adapted:

-companies will face additional costs (e.g. legal information costs, increased insurance costs) (4 agree while 3 disagree),

-insurers will increase risk-premiums due to a lack of predictability of liability exposures (3 agree and 2 disagree) and

-there will be a negative impact on the roll-out of AI technologies in the internal market (3 agree and 2 disagree).

(d) Legal fragmentation

A majority of overall respondents expect that legal fragmentation with respect to civil liability for AI and/or the lack of adaptation of the existing liability framework will, for instance, entail at least to some extent additional costs for companies (e.g. legal information costs, insurance costs) (75 % or 153 out of 205) and the need to adapt AI technologies, distribution models and cost-management models (70 % or 146 out of 205). They will also cause at least some companies to limit their cross-border activities related to the production, distribution or use of AI-enabled products or services (67 % or 135 out of 203).

Regarding business stakeholders in particular, more than half of these respondents expect that, if Member States adapt their liability rules to AI in a divergent way, this will entail at least some additional costs for companies (54%, or 46 out of 85 agree), may reduce cross-border activities involving AI (36% (31 out of 85) ) and have a negative impact on the roll-out of AI technologies (39% or 33 out of 85) while 29% (or 25 out of 85) do not expect such an impact. ).

2.2. Policy options

(a) Policy Option 1

EU citizens, consumer organisations, academic/research institutions and NGOs overwhelmingly supported the measures envisaged under Option 1.

Replies from business stakeholders (business associations and companies/business organisations) were more nuanced: Their agreement and disagreement is roughly evenly distributed as regards rules on the disclosure of information (33%, or 28 out of 86, agreement v. 31%, or 27 out of 86, disagreement). A majority of business stakeholders is against inferring facts from the refusal to disclose information (63% or 54 out of 86, disagreement), shifting the burden of proof (63% or 54 out of 86, disagreement), or presuming causality if the provider of an AI system did not comply with their obligations under the AI Act (41%, or 36 out of 86, disagreement v. 30%, or 26 out of 86, agreement). By contrast, business stakeholders’ views were roughly evenly split about applying such a presumption vis-à-vis the user of the AI-system: while ca. 35 % (or 30 out of 86) of them disagreed with that approach, close to 30 % (or 25 out of 86) supported it.

SME stakeholders were more supportive of measures to ease the burden of proof: The responding individual SMEs approved of those measures. The views of business associations representing (primarily) SMEs were in most cases roughly evenly split.

(b) Policy Option 2

The results of the open public consultation showed support amongst EU citizens, NGOs, academic/research institutions and consumer organisations for either full or minimum harmonisation of strict liability for the operation of certain AI-enabled products and the provision of certain AI-enabled services. In particular, EU citizens favoured a ‘full harmonisation’ approach to strict liability (77 %, or 68 out of 88, v. 54 %, or 47 out of 88, supporting minimum harmonisation 224 ). By contrast, more responding consumer organisations supported a minimum harmonisation approach, allowing MS to maintain broader and/or more far-reaching national strict liability schemes (4 out of 7 agreed with a minimum harmonisation of strict liability). None of the responding consumer organisations supported full harmonisation of strict liability.

Business stakeholders tend to oppose the harmonisation of strict liability. Opposition was stronger regarding a minimum harmonisation approach (70 %, or 59 out of 84, v. 14%, or 12 out of 84, support) than a full harmonisation of strict liability (42 %, or 35 out of 84, disagreement v. 30 %, or 28 out of 84, agreement). Interestingly, SME stakeholders were distinctly more supportive of a harmonised strict liability regime: Almost all of responding individual SMEs (9) approved of this policy option, and business associations representing (primarily) SMEs were evenly split regarding both the minimum and full harmonisation of strict liability. Business stakeholders consistently argued against the harmonisation of strict liability in their position papers.

(c) Policy Option 3

The ‘staged approach’ (= preferred option) was not presented as a separate policy option in the public consultation, because it is a combination of elements included under the other policy options. Therefore, there are no results linked specifically to this PO. The staged approach was developed and refined in light of feedback received from stakeholders throughout the impact assessment process. It strikes a balance between the needs expressed and concerns raised by all relevant stakeholder groups.

3. Detailed summary of results

The following summary explains for each of the topics covered by the public consultation the respective results as regards the problem analysis as well as the development and assessment of policy options and how they were taken into account in the IA.

3.1.    Problems to be addressed: Lack of compensation and consumer trust, legal uncertainty and fragmentation

(a)    Problem for victims of damage caused by AI: Lack of compensation

The problem of lack of compensation due to the specific characteristics of AI was largely confirmed.

-74% (or 167 out of 227) of overall respondents (56%, or 127, strongly) agreed that in the case of AI that lacks transparency and explainability, it could be difficult for injured parties to prove that the conditions of liability (fault or causation) are fulfilled. Only 12 % (or 26 out of 227) of respondents disagreed.

Looking at specific stakeholder groups, all responding consumer organisations and national ministries as wells as research institutions confirmed this problem. An overwhelming majority of EU citizens (91% or 85 out of 93) also agreed. Also a relative majority of businesses, including business associations and individual companies, agreed with this statement (41% or 35 out of 85), while 28% (or 24 out of 85) disagreed.

-On the question about whether it could be difficult to link damage caused by AIto the actions or omissions of a human actor when AI operates with a high degree of autonomy, approximately 67% of respondents agreed (153 out of 227, of which 105 strongly, i.e. ca. 46%) that it could be difficult to link damage caused by highly autonomous AI to the actions or omissions of a human actor. 20% (or 47 out of 227) disagreed. In this respect, high agreement was reached amongst responding citizens (84% or 78 out of 93) and almost unanimous agreement (94%, or 31 out of 33, with only one disagreeing) amongst responding consumer organisations, national ministries and academic/research associations. The same applies for the responding national ministries and research institutions. Businesses were split on this question: while 36% (or 30 out of 85) agreed with this question, 45% (or 38 out of 85) disagreed.

-61 % (or 139 out of 232) of respondents also agreed (42%, or 95 out of 232, strongly) that because of specific AI characteristics, victims of damage caused by AI may in certain cases be less protected than victims of damage that did not involve AI; only 23% (or 52 out of 232) disagreed. While a majority of businesses disagreed with such statement (54% or 45 out of 85), 26% (or 22 out of 85) did think that victims will be less protected due to the specific characteristics of AI. Respondent consumer organisations, national ministries and EU citizens overwhelmingly confirmed this problem.

The positions submitted in writing point to a similar trend: Representatives of academia, NGOs and consumer associations agree that AI characteristics make it more difficult for victims to claim compensation and some agree with the legal uncertainty of applying existing rules. Many business associations and companies cast doubt on the specific challenges linked to AI, although some admit that victims might have difficulties claiming compensation.

(b)    Problem of lack of trust in AI – reduced uptake

For a majority of respondents the lack of adaptation of the current liability framework to AI may negatively affect trust in AI (60%, or 135 out of 227, agreed; 25%, or 58 out of 227, disagreed) and the uptake of AI-enabled products and services (56%, or 126 out of 227, agreement v. 27%, or 60 out of 227, disagreement).

NGOs and consumer associations overwhelmingly confirm these problems (89%, or 16 out of 18, expect a lack of trust and 78%, or 14 out of 18, reduced uptake). Similarly, a large majority of EU citizens consider that the lack of adaptation of the current liability framework may negatively affect trust in AI (81% or 77 out of 95) and the uptake of AI-enabled products and services (79% or 74 out of 94). Likewise, the vast majority of research institutions share this opinion. Among responding national ministries 1 confirmed the problem and 1 denied it. While the majority of business stakeholders (business associations and individual companies) (58% or 49 out of 85), did not confirm these particular problems, significant shares of these stakeholders did consider that current liability rules will lead to lower consumer trust (27% or 23 out of 85) and uptake (25 % or 21 out of 85), if not adapted. Moreover, amongst those business stakeholders who considered that AI can make it difficult to claim compensation:

-64% agreed that the lack of adaptation of the current liability framework to AI may negatively affect trust in AI, and

-59% considered that the lack of adaptation of the current liability framework to AI may negatively affect the uptake of AI-enabled products and services.

(c)    Legal uncertainty

Almost 63% (or 144 out of 227) of respondents agreed (36%, or 81 out of 227, strongly) that there is uncertainty as to whether and how liability rules under national law apply to damage caused by AI; only 22% (or 50 out of 227) disagreed (only 5 (11) strongly).

Consumer associations unanimously confirmed their perception of legal uncertainty, and responding EU citizens NGOs and research institutions agreed with a very large majority. In particular, 85% of EU citizens (or 80 out of 94, with only one citizen disagreeing) and 84% of responding NGOs, consumer organisations and academic/research institutions (or 27 out of 32, with none disagreeing) agreed that it is uncertain whether and how liability rules under national law apply to damage caused by AI. Responding national ministries were split. While about half of businesses stakeholders (52% or 44 out of 85) do not perceive legal uncertainty specifically with respect to liability for AI, about one third (32% or 27 out of 85) of these respondents did confirm that problem as well.

This picture is mirrored by the responses to the question whether it is uncertain how national courts will address possible difficulties of proof and liability gaps in relation to AI. An overwhelming majority of responding NGOs + consumer organisations (77% or 13 out of 17) agreement; 0% disagreement) and EU citizens (91% or 84 out of 92, agreement) agreed that the approach of national judges is uncertain (while some of these respondents did not have an opinion on this, none disagreed). 2 out of 5 responding national ministries also confirmed the existence of such uncertainty, while one disagreed. Respondents from the business side were again evenly split on this: 34% (or 29 out of 85) confirmed the existence of such uncertainty while 43% (36 out of 85) did not.

Businesses stakeholders in particular were asked to provide feedback on the expected consequences of a lack of adaptation of the existing liability rules to AI. In this respect:

-Substantial shares of businesses stakeholders (although not the overall majority) think that the lack of adaptation of the current liability rules to AI will entail a number of negative economic consequences. They expect for instance:

othat this lack of adaptation of liability rules will entail additional costs (e.g. legal information costs, insurance costs) for companies (27% or 23 out of 85),

othat companies may defer or abandon certain investments in AI technologies if the current liability framework is not adapted (20% or 17 out of 85) and refrain from using AI when automating certain processes (21% or 18 out of 85).

-Interestingly, when looking specifically at business associations representing primarily SMEs, opinions are even more balanced. In particular, a slight majority of these respondents slightly agree with the statements that companies will face additional costs (e.g. legal information costs, increased insurance costs) (4 agree while 3 disagree), that insurers will increase risk-premiums due to a lack of predictability of liability exposures (3 agree and 2 disagree) and that if the liability framework is not adapted there will be a negative impact on the roll-out of AI technologies in the internal market (3 agree and 2 disagree). They slightly disagree with the statement that if the framework is not adapted companies may refrain from using AI when automating certain processes (4 against 3) and that there will be higher prices of AI-enabled products and services (4 against 3).

-Most of the few responding individual SMEs expect additional costs, reduced investments and cross-border business activities, higher prices and insurance premiums as well as a negative impact on the roll-out of AI as a result of the lack of adaptation of current liability rules.

-When focusing specifically on those business stakeholders who agreed that there is legal uncertainty regarding the application of national liability rules to damage caused by AI, strong majorities expect that, if the current liability framework is not adapted:

ocompanies will, at least to a moderate extent, face additional costs (82%);

ocompanies may refrain, to a moderate, large or very large extent, from using AI when automating certain processes (66 %);

ocompanies may limit their cross-border activities related to the production, distribution or use of AI-enabled products or services (62%);

oinsurers will increase risk premiums, at least to a moderate extent, due to a lack of predictability of liability exposures (67%).

(b)    Legal fragmentation

If Member States adapt liability rules for AI in a divergent way, or national courts follow diverging interpretations of existing liability rules in the case of damage caused by AI, 75% (or 153 out of 205) of respondents expect that this will entail at least some additional costs for companies active in AI (e.g. legal information costs, increased insurance costs). 70% (or 146 out of 205) also expect legal fragmentation with respect to liability for AI to entail a need to adapt AI technologies, distribution models and cost management models at least to some extent.

Businesses stakeholders in particular were asked to provide feedback on the expected impacts of AI-specific legal fragmentation. Amongst these respondents:

-A majority (54% or 46 out of 85) expects that legal fragmentation with respect to civil liability for AI will entail at least some additional costs for companies (e.g. legal information costs, increased insurance costs). Only 15% (or 13 out of 85) of businesses would not expect such costs at all.

-A relative majority (46% or 39 out of 85) also expects that legal fragmentation with respect to liability for AI entails a need to adapt AI technologies, distribution models and cost management models, while only 18% (or 15 out of 85) disagree. 59% (or 49 out of 83) of responding businesses stakeholders anticipate that fragmented AI-specific liability rules would entail at least some (including small) need for technological adaptations when providing AI-based cross-border services (v. 27% or 22 out of 83 who do not anticipate such a need for adaptation at all).

-50% (or 42 out of 84) expect that insurers will increase premiums due to more divergent liability exposures (whereas 14% or 12 out of 84 do not).

-36% (or 31 out of 85) expect that companies may limit their cross-border activities related to the production, distribution or use of AI-enabled products or services, while only 28% (or 24 out of 85) would not.

-Similarly, 38% (or 32 out of 85) think that legal fragmentation regarding liability for AI will lead to higher prices for AI-enabled products and services (while only 24%, or 21 out of 85, disagree).

-39% (or 33 out of 85) expect that increased legal fragmentation regarding liability for AI would have a negative impact on the roll-out of AI technologies, whereas only 29% (or 25 out of 85) dispute this impact.

Responding business associations representing in particular SMEs tend to agree even more strongly that legal fragmentation would lead to additional costs, reduced investments and cross-border activities, higher prices and insurance premiums, as well as a negative impact on the roll-out of AI. For instance, 7 of these respondents agreed that legal fragmentation can entail additional costs for companies (e.g. legal information costs, increased insurance costs) when producing, distributing or using AI-enabled products or services, while only 1 of them disputed this. The same trend was confirmed by the responding individual SMEs.

3.2.    Need for EU action

The public consultation results showed clear overall support for EU action on AI liability. When asked to rank policy options by order of preference from 1 (most preferred) to 8 (least preferred), respondents gave by far the lowest average score to the option ‘No EU action’ (baseline). 225 There was particularly strong support for EU action amongst consumer organisations and NGOs, almost all of which chose ‘No EU action’ as their least preferred option. Likewise, EU citizens clearly support EU action. 226

The following chart shows that:

-only 14% of respondents supported No EU Action;

-only 11% expressed a preference for addressing AI-specific issues exclusively by easing the burden of proof under the PLD;

-75% of respondents preferred one of the policy options involving harmonised measures to adapt also national liability rules to AI:

While support for EU action on AI liability was consistently strong across responding EU citizens, consumer organisations, NGO and research institutes, it was a bit less pronounced amongst business stakeholders in particular. Nevertheless, as shown by the following graph, while a preference for ‘No EU Action’ was considerably more widespread amongst these respondents (32 %), the largest share nevertheless expressed a preference for one of the policy options involving harmonised measures to adapt also national liability rules to AI:

3.3. Policy options

(a) AI-specific measures to ease the burden of proof under national civil liability rules

The public consultation showed overall support for targeted EU measures to ease the victim’s burden of proof under existing national liability rules:

-More than two third of respondents (67%) agreed that the defendant (e.g. producer, user, service provider, operator) should be obliged to disclose necessary technical information (e.g. log data) to the injured party to enable the latter to prove the conditions of the claim (v. only 14% who disagreed with this policy option). Consumer associations and research institutions unanimously supported this option and EU citizens were also strongly in favour of it. Amongst business stakeholders, a slightly bigger share (33%, or 28 out of 86) agreed with this policy approach than disagreed (31% or 27 out of 86).

-In connection with the previous question, 57% of respondents agreed that if the defendant refuses to disclose relevant information, courts should infer that the conditions to be proven by that information are fulfilled (v. 33% who disagreed with this). This option received strong support from responding NGOs + consumer organisations (61% support v. 17% disagreement) and EU citizens (84%, or 77 out of 92) support v. 13%, or 12 out of 92, disagreement). By contrast, business stakeholders were more sceptical in this regard: close to two thirds (63% or 54 out of 86) of these respondents were opposed to such a rule.

-51% (or 114 out of 225) of respondents consider that if the provider of an AI system failed to comply with their safety or other legal obligations to prevent harm (e.g. those proposed under the proposed AI Act), courts should infer that the damage was caused due to that person’s fault. While all the other stakeholder categories support this approach, with consumer associations supporting it unanimously, businesses were more split: 30%, or 26 out of 86, agreed with his option and 41%, or 36 out of 86, disagreed.

-The option to apply a similar mechanism (presuming causality in the case of non-compliance with AI Act) to users of AI systems, received more support from businesses (35%, or 30 out of 86, agreed, while 29%, or 25 out of 86, disagreed) and was supported also by a clear majority of the other stakeholder categories.

-63 % (or 142 out of 226) of overall respondents consider that if, in a given case, it is necessary to establish how a complex and/or opaque AI system (i.e. an AI system with limited transparency and explainability) operates in order to substantiate a claim, the burden of proof should be shifted from the victim to the defendant in that respect (v. 27% (or 61 out of 226) who disagreed with this option). While about a quarter of business stakeholders also supported this approach, a clear majority (63% or 54 out of 86) disagreed with it.

The position papers submitted in the consultation revealed similar trends: representatives of consumer organisations, citizens and NGOs agree that victims should not bear the burden of proof and there is a need for an alleviation while opinions vary about how this could be done (for example some consider that access to information is not helpful as the victim would not be able to understand and use it). Business organisations and companies usually do not favour a shift of the burden of proof or state that sufficient alleviations can be granted by courts under the various national laws. In addition, some of these stakeholders caution against an obligation to make technical information available to the victim, because of intellectual property considerations. Academia recognises the difficulties in finding the right balance and is suggesting, for example, to focus on further defining duties of care or introducing rebuttable presumptions for breaches of such duties of care.

(b) Strict liability and mandatory insurance

EU citizens, NGOs and research institutions were strongly in favour of harmonising strict liability for operating AI-enabled products and providing AI-enabled services, limited to cases where these activities pose serious injury risks to the public:

-EU citizens: 54% (or 47 out of 88) agreement with minimum harmonisation of strict liability; 77% (or 68 out of 88) agreement with full harmonisation;

-NGOs + research institutions: 73% (19 out of 26) agreement with minimum harmonisation of strict liability; 69% (18 out of 26) agreement with full harmonisation.

They also tended to agree with laying down a harmonised insurance obligation, where it does not exist yet, covering such a strict liability regime.

Consumer organisations supported strict liability only in the form of minimum harmonisation, allowing Member States to maintain broader and/or more far-reaching national strict liability schemes applicable to other AI-enabled products and services (4 agreed with minimum harmonisation of strict liability v. 3 that disagreed; 6 disagreed with full harmonisation whereas none agreed with this approach).

Written contributions submitted by citizens, consumer organisations, NGOs and research institutions were not very clear on the strict liability option; the ones that supported it did not provide much input or referred to the need to exempt consumers from strict liability, to take risks and circumstances of use into account or to allow national law to have in place more favourable regimes.

The survey results showed that business stakeholders tend to disagree with the harmonisation of strict liability and mandatory insurance, but their disagreement was stronger as regards a minimum harmonisation approach to strict liability (70% opposition (59 out of 84) v. 14% support (12 out of 84) than a full harmonisation approach (42% disagreement (35 out of 84) v. 30% agreement (28 out of 84)). In the position papers, business associations and companies almost universally opposed a harmonisation of strict liability, for various reasons: some consider that AI does not warrant strict liability in general as not being more dangerous than other technologies, others consider that it is too soon, as uses of AI that would pose serious risks are not yet here and harmonising it would deter innovation, while others believe that existing national strict liability regimes are sufficient. The few representatives of this category that supported strict liability had done so to the extent it would be clearly limited to a few high-risk cases, if the heads of claims would be restricted or if clear exceptions would also be harmonised.

(c) Types of compensable harm

Stakeholders were asked whether an EU instrument on AI liability should harmonise the types of harm that give rise to a compensation claim when caused by AI. Here again, views varied widely between different stakeholder categories.

Amongst EU citizens, consumer associations, NGOs and research institutions there was clear support for EU rules prescribing the compensability of pure economic loss (e.g. loss of profit), loss of or damage to data (not covered by the GDPR) and immaterial harm (like pain and suffering, reputational damage or psychological harm).

By contrast, business stakeholders were mainly opposed to such rules. Their opposition was strongest regarding the compensability of immaterial harm (63% (or 52 out of 83) disagreement v. 23% (or 19 out of 83) support) and of data loss/damage not resulting in a verifiable economic loss (75% (or 63 out of 84) opposition v. 12% (or 10 out of 84) support). This picture is confirmed also by written contributions submitted by business stakeholders. They namely emphasised the very diverse approaches under existing national rules and the concern that harmonisation for AI-related damage would create compares to cases not involving AI.

C.    Other consultation activities

1. Public consultation launched by the White Paper on AI and the Commission report on safety and liability

An open 16 weeks web-based public consultation ran from February 19 to June 14, 2020. The objective of the consultation was to collect views/opinions on the White Paper on Artificial Intelligence - A European Approach to excellence and trust, COM(2020) 65 and on the Commission Report on the safety and liability implications of Artificial Intelligence, the Internet of Things and robotics, COM (2020) 64. The public consultation resulted in more than 1200 responses from all categories of stakeholders from across the EU.

63 % of respondents to the open public consultation on the AI White Paper in 2020 were in favour of adapting national liability rules for all or for specific AI applications to better ensure proper compensation and a fair allocation of liability. In that context, citizens were particularly supportive with 72% of respondents calling for some adaptations. While the overall support among companies was 45%, it was higher among SMEs with 60% of respondents.

2. Online interactive webinars

Twelve webinars were organised from 8 to 12 June 2020 for a total of 18 hours of interactive discussions with stakeholders. The webinars had the objective to inform the Commission as part of the consultation launched by the AI White Paper and to encourage stakeholders to participate in the public consultation.

211 stakeholders registered and 170 participated. Many stakeholders’ categories were represented: the majority were from the tech industry, including medical tech companies, producers and software developers. Organisations representing consumers were also well represented as well as academia and legal professions. Interested citizens and non-European stakeholders also participated. A successful effort in the communication strategy was put in place to make sure that the most important organisations representing SMEs and start-ups participated in the discussions.

The webinars covered three main topics related to liability: how to define high-risk AI applications, who should be held strictly liable and possible measures to help victims in case of fault-based rules.

2.1. Webinar 1: Risk-based approach: How should high-risk and non-high-risk AI applications be determined?

Stakeholders provided feedback on the risk based-approach presented in the AI White Paper and on the possible criteria to identify the AI applications with a specific risk profile. Several stakeholders asked for consistency on what is high risk in the different dimensions presented in the WP (fundamental rights, safety and liability). However, a slight differentiation could be justified with respect to strict liability, which should have a narrower scope.

A European organisation representing consumers stressed that consumers should be entitled to an effective redress mechanism and access to justice irrespective of the risk level. Another consumer organisation stated its preference for a precautionary approach taking into account how new an application is and how little it is known about the risks it can raise. Interests of minorities should also be considered.

A number of participants emphasises that a possible liability instrument would need to fit into the current framework. The need for a mandatory insurance should also be carefully considered. An organisation representing start-ups added that strict liability for high-risk AI applications could stifle innovation. The same representative also stated that any legislation put forward needs to be easy for companies to understand. Compliance is difficult for many start-ups. Other participants emphasised that any policy measures should be based on empirically established issues to address real gaps. Some stakeholders representing different interests argued that the level of autonomy should be considered for the criteria to define AI applications with a particular risk profile. A greater degree of autonomy leads to a higher risk. Another element to consider is whether the AI system merely supports the human decision or replaces it. While participants largely agree that the AI currently on the market is merely assisting human decisions, it was highlighted that some challenges might arise in those cases where AI is autonomous and continues to develop once it is put on the market. The AI could for example learn to act outside previously awarded safety certificates.

2.2. Webinar 2: Strict liability of the operator of high-risk AI applications: Who is the operator in that case?

The majority of the speakers agreed in principle that the liable person could be defined based on the criterion of control. Some stakeholders representing tech companies mentioned that the criterion of “economic benefit” should also be considered, particularly in the B2B context. In addition, it was stated that the person who is closest to the operations would probably be the best addressee for liability. Some stakeholders were also in favour of joint and several responsibility, with the possibility to take redress along the supply chain. A number of stakeholders representing different interests were of the opinion that in the case of strict liability there is a need for mandatory insurance and caps are important, otherwise it would be difficult for the insurers to calculate the coverage. Medtech representatives supported caps in the medical industry. Mandatory insurance should be without prejudice to developers’ claim against the players in the upstream value chain. Stakeholders representing the automobile industry voiced scepticism regarding the idea of regulating liability caps at EU level because Member States do not calculate damage the same way. A lawyers’ association brought the counter-argument that different caps in different MS could result in forum shopping. Representatives from industry and legal professions took the view, with some variations, that exceptions/defences are important, namely:

-Force majeure - Several speakers were of the opinion that an exception from liability in unforeseen circumstances should be applicable

-Contributory damage – most speakers were also in favour of contributory damage (e.g. in cases where the victim doesn’t carry out a necessary update);

-Cybersecurity attacks/3rd party interference – several speakers were in favour of such an exception. However, some industry representatives argued against an absolute exception, but in favour of a defence allowing the liable person to demonstrate that he or she has complied with cyber standards;

-Open source AI – Some representatives from academia suggested that the liability of the developer of open source AI should be limited;

-A few stakeholders mentioned the misuse of AI by a user.

By contrast, representatives of consumer interests spoke against any kind of defence, caps, threshold or other exceptions, because digital goods expose consumers to particular difficulties when it comes to claiming compensation.

2.3. Webinar 3: Burden of proof for AI applications other than those with a specific risk profile: How can it be made easier for victims of damage caused by such applications to prove their claim?

Representatives of consumer interests emphasised the crucial importance of the burden of proof for all AI products, stressing that it is it is almost impossible for the consumer to prove fault and even harder to prove causality. Consumers should have to prove only the existence of a damage, because the prevailing information asymmetry between consumers and professionals justifies a reversal of the burden of proof with respect to all other liability conditions. These arguments were seconded by some representatives from academia, who stated that the difficulty of proving fault is particularly relevant in discrimination cases. An organisation representing businesses, in particular SMEs, added that a reversal of the burden of proof was necessary not only for consumers but also for companies that suffer harm.

By contrast, several representatives of the tech and software industry argued against a reversal of the burden of proof for AI, questioning the need to deviate from the general rules. One of these discussants stressed that providers of AI-services or products should not be required to prove a negative fact (i.e. that they did not commit a faulty act). Large software companies argued that a reversal of the burden of proof was a very sharp tool and represented a significant burden for the defendant. Therefore, such a tool should only be considered on a case-by-case basis if it is prohibitively difficult for the victim to prove a claim. Representatives of tech companies submitted that the claimant should in any case be required to provide at least prima facie evidence that the defendant did something wrong. Representatives of manufacturing companies stated that a reversal of the burden of proof should apply only under the condition that it is impossible to ascertain the facts of the case, for instance in the case of traffic accidents. Legislative modifications of the burden of proof should be limited in scope based on the risk and the need for protection.

3. Feedback on inception impact assessment

Thirty-four stakeholders submitted feedback throughout the feedback period (30th June 2021 – 28th July 2021). Many categories were represented. The majority of the stakeholders came from the tech industry, including medical tech companies and software developers. Several NGOs submitted feedback, as well as some organisations representing consumers. EU citizens, public authorities and academic institutions also participated.

Some stakeholders support a minimum harmonisation or an approach based on non-binding measures.

A number of stakeholders urged for a risk-based approach, stressing that AI applications ought to be treated according to the specific risk each of them poses.

In order to provide legal certainty and avoid fragmentation among MS, several stakeholders from different sectors favour a broader harmonisation of rules concerning extra-contractual liability. Some industry stakeholders argued that the producer should not be burdened with strict liability for damage caused by AI, as it would be impossible for him/her to anticipate every conceivable use of the product or service. This is because AI applications can be employed in unlimited scenarios and can even be misused or used in ways not specified by or foreseeable for the producer.

Many stakeholders from NGOs and consumers’ organisations expressed support for an alleviation of the burden of proof, in order to overcome challenges typical of AI which make it difficult for the injured parties to get compensation (such as autonomous behaviour, continuous adaptation, limited predictability and opacity). Another aspect that makes the burden of proof particularly heavy for the injured parties is their lack of technical information. Within the tech industry, while some stakeholders are generally in favour of a promotion of the exchange of technical information throughout the value chain to ensure that AI systems remain predictable, others cautioned that no obligation to disclose technical information should be imposed, if it infringes the intellectual property or legitimate confidentiality interests of the producer. An EU citizen expressed support for the introduction of an obligation to disclose technical information, whereas an NGO argued that such an obligation, although generally desirable, would not be sufficient to educate consumers about whether or not the AI system is defective in their case.

Several stakeholders support a reversal of the burden of proof, whereas the majority of industry representatives opposed it, arguing that it would deter innovation. However, an NGO counter-argued that such a reversal would incentivise producers to invest more in preventing defects, which would in turn increase the customers’ willingness to pay higher prices for products that they reasonably trust to be safe. Finally, some companies caution against a blanket reversal of the burden of proof, since such a tool should only be applied to very specific cases as a measure of last resort.

Finally, medtech companies maintained that, whenever AI is employed in the healthcare system, it is only used as a supporting tool, while the final decision rests with the healthcare professional who should then ultimately be responsible for his/her choices.

4. Surveys and targeted interviews carried out in the framework of two supporting studies (general economic and behavioural analysis)

Two external studies commissioned to inform this IA contributed to the consultation of stakeholders:

-an economic analysis of AI-specific problems affecting current civil liability rules 227 ;

-a behavioural economics study investigating the link between liability-related problems and policy measures and societal trust in / consumer uptake of AI-enabled products and services 228 .

The methodology implemented by these studies is described in more detail in Annex 4. The following summary is limited to the tasks relevant as part of the consultation activities for this IA.

4.1. Economic study

-Consultation activities to collect data on the costs and time required to claim compensation under current liability rules, as a means of approximating victims’ difficulties in claiming compensation due to the specific characteristics of AI: The study team consulted legal experts in 13 Member States and the UK, with long-standing experience in civil liability cases related to information and communication technologies and novel technologies. The legal experts estimated the costs and time required by lawyers and technical experts in “traditional” cases of damage, as well as in cases involving AI applications. The data collection was complemented by desk research. The results confirmed that AI can make it substantially more difficult for victims to claim compensation (see Annex 4 for more detailed explanations).

-Consultation activities to collect data on the impact of liability-related legal uncertainty and fragmentation on the roll-out of AI: The study team principally used two methods to collect data for this task: (i) A survey targeted to European trade associations and companies with activities linked to AI, and in particular the operation, production, development of AI systems in relation to the six selected use cases; (ii) Interviews with trade associations, companies and experts. The interviews were conducted in a semi-structured manner, with an overall list of questions shared in advance and the possibility to address also different topics during the interviews. The findings supported the consultant’s overall conclusion that legal uncertainty and fragmentation regarding civil liability for AI entail significant internal market barriers

-Consultation activities to assess the impact of preliminary policy options: Here as well, the study team used a combination of survey and interviews. In order to receive targeted data across different relevant stakeholder groups, the survey was designed in three different versions: (i) for operators of AI-enabled products and providers of AI-enabled services (including companies that are at the same time manufacturers/developers of AI-systems); (ii) for the legal experts consulted also on the cost of claiming compensation under the current liability rules, to ascertain whether the preliminary policy options were suitable to improve victims’ situation; (iii) for insurance companies. The results supported the consultant’s conclusion that policy measures to ease the burden of proof, in combination with a limited harmonised strict liability regime for AI, would significantly increase legal certainty, prevent fragmentation, and thus have a significant positive effect on the functioning of the internal market in AI-enabled products and services.

-The study team organised and hosted a webinar during the month of October 2020. The webinar enabled over 40 participants from various industrial sectors to provide views on the preliminary findings of the study and to debate the major research questions (impact of legal uncertainty, future legal fragmentation, and preliminary policy options). The webinar allowed the study team to confirm and adjust the findings and provided useful input to orientate the subsequent data collection activities.

4.2. Behavioural economics study

In the framework of the behavioural economics study, the consultant sought stakeholder feedback was through the following activities:

-In-depth interviews with consumers, which helped to deepen the understanding of the behavioural dynamics that link, on the one hand, societal acceptance of AI applications, consumers’ trust and their willingness to take up AI applications and, on the other hand, civil liability rules. The insights derived allowed the consultant to fine-tune the design of the survey and of the survey-based experiment, while also partly complementing their results. The interviews were carried out individually with 18 consumers from three countries.

-An online survey carried out on representative samples of the adult population from eight countries (Denmark, Netherlands, Ireland, France, Germany, Italy, Poland, and Romania), with a sample size of at least 1,000 respondents per country, amounting to a total of 8,079 respondents across the eight countries. The sample for each country was a quota-based one, with additional demographic weights being computed and appended to the data in order to correct any potential distortions of the characteristics of the respondents within the sample from the characteristics of the population from which the sample was extracted.

The results of the behavioural survey and experiments showed that the perception of liability rules is one of the factors shaping societal trust and consumers’ willingness to take up AI-enabled products and services (see Annex 4 for a more detailed summary of findings).

5. Workshop with Member States

A workshop with representatives of MS’ governments was organised on 1 February 2022, with a view to gathering MS’ views and input on the problem definition and policy options before finalising this IA. Most MS did not express a clear position yet, but some conveyed preliminary observations.

-Need for EU action: Some MS formulated reservations regarding the need for harmonisation of liability for AI, considering notably that existing national liability rules are sufficiently flexible to deal with the specific challenges of AI.

-Types of AI to be covered: Several MS favoured a narrow definition of AI, e.g. focused on highly autonomous and/or opaque AI.

Fault-based liability: A few MS were sceptical about harmonised measures to ease the burden of proof under fault-based liability rules, or at least encourage a cautious approach. However, other MS were open to the adaptation of rules on access to information to help victims make a successful claim as well as the targeted adaptation of the burden of proof.

-Some MS expressed the view that the types of damages compensable for harm caused by AI ought to be determined at national level or that the EU proposal should only cover economic loss, whereas the rest should be left to national law. Some MS oppose the harmonisation of rules regarding contractual limitations or exclusions of liability while one MS would welcome such intervention at EU level.

-Strict liability: One MS deems it advisable to reassess the need for strict liability after the roll-out of AI-enabled products and services. According to another MS, if strict liability is considered, then a risk-based approach should be followed, although the definition of high risk is yet to be determined. Two MS opposed a harmonisation of strict liability in their responses to the public consultation. Some MS stated at the workshop that the initiative should not lower the level of protection provided by national law and that B2B, C2C and B2C relationships ought to be treated equally.

-Insurance: MS that provided input are of the opinion that, should strict liability be introduced, mandatory insurance should also be considered. However, it should first be evaluated whether the insurance sector offers such coverage on the market. A few additional points were raised: such obligation ought to be enshrined within the relevant national legal acts; mandatory insurance might cause an increase of premiums and it should be technology-neutral to ensure proportionality.

Annex 3: Who is affected and how?

1.Practical implications of the initiative

1.1.Implications for the most relevant stakeholder groups

The preferred option has two elements: targeted alleviations of the burden of proof and a targeted review mechanism to re-assess the need for more far-reaching measures, including strict liability and possibly a mandatory insurance for the use of certain AI-technologies. Given that the review mechanism would not have practical implications for stakeholders, as it might lead only in the future to changes of the proposed Directive, this analysis will focus on the implications of the targeted alleviations of the burden of proof.

(a) Impacts on businesses (in particular as potentially liable parties):

The implications for businesses would mainly concern an increased legal certainty and a reduced fragmentation. It will be easier for companies to estimate liability risks and the related costs. While under the baseline scenario courts might apply on an ad-hoc basis alleviations of the burden of proof to remedy what they consider an unequitable result, clear and harmonised alleviations of the burden of proof will help companies know what to expect in case AI is involved both domestically and cross-border. Although for what concerns cross-border relations Member States can impose more stringent measures (for instance strict liability), companies would still benefit from reduced compliance costs compared to the very fragmented baseline scenario. Such clarity might also help companies get more appropriately priced liability insurance coverage. The combined economic gains linked to these impacts have been approximated in terms of increases in the AI market size. For further explanations on this quantification of indirect economic benefits, see heading 1.2.(a) below and the cost-benefits table at the end of this Annex.

These benefits outweigh the adjustments costs as the business-as-usual costs under the baseline scenario related to the uncertainty to assess what liability rule would apply to AI and what burden-of-proof rule a court would apply in a concrete case are higher. These benefits also outweigh the redistribution effects linked to the alleviations of the burden of proof: an improved liability regime would allow shifting the costs of the damage born by the victim to the wrongdoer. However, this shift cannot be considered as an undesirable impact or undue burden. Firstly, it is in line with one of the fundamental justice-related purposes of liability law, i.e. to ensure that a person who harms another person in an illegal way will compensate the harm caused to the victim. Secondly, it is in line with the Commission policy objective to ensure that victims of damage caused with the involvement of AI systems have the same level of protection as victims of damage caused by other technologies. Finally, it also achieves a more efficient cost-allocation, i.e. to the person having caused the damage/best placed to prevent damage from occurring, instead of being borne by the victim.

(b) Impacts on victims of damage caused by AI (natural persons and businesses) 

The targeted alleviations of the burden of proof would allow victims of damage caused by AI to have the same level of protection as victims of damage caused by other technologies. This means that it would be prevented that victims of AI induced damage would be left without compensation. Instead, the economic cost for compensating the harm occurred would be allocated to the person who is indeed responsible for causing the damage. In addition, victims’ costs linked to the burden of proof would be reduced, including the costs of expert analysis, by ensuring access to relevant information and alleviating the victim’s burden of establishing how or why an AI-system arrived at a certain output.

(c) Consumers

The practical benefits for businesses would also have impacts on consumers because of the faster rollout of AI-technologies. In particular, consumers will benefit from faster and more personalised services, innovative and performant products as well as advances in the fields of health, safety, security, mobility, sustainability, circular economy, media, etc.

(d) Insurance companies

Targeted harmonisation and adaptation of the rules on the burden of proof may lead to a slight increase in the take-up of insurance policies by liable parties due to increased legal certainty and reduced fragmentation. These aspects increase awareness of liability risks and create more favourable conditions to offer insurance policies.

1.2. Broader economic and societal implications

(a) Implications for the European economy

The targeted and harmonised alleviations of the burden of proof would contribute to an efficient cost allocation. By increasing legal certainly and reducing fragmentation, the harmonised rules would incentivise the rollout of AI and consumer trust. Its combined impacts are expected to have a positive effect on cross-border trade in AI-enabled products and services and the development of the European AI-sector as a whole. 229 The economic study commissioned for this impact assessment estimated that a combination of alleviations of the burden of proof and measures to harmonise strict liability for certain AI-enabled products and services would increase the cross-border trade in the AI-enabled goods and services falling under the six use-cases analysed in depth for that study by about 5 %. While the preferred option does not include all of the assumptions made for that estimation (the need for strict liability would be assessed at a later stage), it is nevertheless relevant because the decisive drivers of the expected economic benefits – increased legal certainty, reduced fragmentation and increased consumer trust – are likely to materialise also under targeted and harmonised alleviations of the burden of proof. Under the same assumptions, the expected increase in cross-border trade could translate to an increase by EUR 150 million of the production value of cross-border trade in those products and services alone. 230 While these figures give some conservative indication of the direction of economic impacts, they do not provide an estimate of their overall magnitude. As they do not take into account trade in products and services within individual MS and only refer to the six specific use-cases, the overall economic impact is likely to be significantly larger, because the overall affected market size is much larger: the preferred policy option is estimated to generate an overall increase in the AI market size by ca. EUR 500 mln (low estimated value) to EUR 1.1 bln (high estimated value) 231 .

(b) Environmental impacts

On a general level, it is expected that AI-solutions can generate efficiencies and contribute to the innovation of environmentally friendly technologies. By contributing to an increased roll-out of AI-enabled products and services, the preferred policy option might hence yield certain indirect environmental benefits. However, it is acknowledged that there is no sufficient basis for assessing such effects even approximatively, because they are too far removed from the envisaged adaptations of liability rules.

(c) Impacts relating to the UN sustainable development goals

By incentivising the development and rollout of AI, the preferred option would support several targets 232 across all the Sustainable Development Goals (SDGs). The Covid pandemic has already showed the potential of AI to contribute to Goal 3 on good health and well-being, as it can contribute to finding effective vaccines, detecting diseases via pattern recognition using medical imagery, calculating probabilities of infection, or providing emergency response with robots replacing humans for high-exposure tasks in hospitals 233 . AI can also be an enabler by supporting the provision of food, health, water and energy services to the population. It can also support low-carbon system by supporting the creation of circular economies and smart cities that efficiently use their resources 234 . Looking at SDGs related to the environment, especially Goal 13 on climate action, there is evidence that AI can support the understanding of climate change and the modelling of its possible impacts. For instance, AI techniques can help to identify desertification trends over large areas, information that is relevant for environmental planning, decision-making and management to avoid further desertification, or help reverse trends by identifying the major drivers 235 .