Rules for trustworthy artificial intelligence in the EU

KEY POINTS

What is an AI system?

An AI system is a machine-based system designed to operate with some level of autonomy that can:

• adapt after it is deployed; and
• generate outputs such as predictions, content, recommendations or decisions from input it receives (to achieve explicit or implicit objectives).

A risk-based approach

The legislation follows a risk-based approach, which means the higher the risk of causing harm to society, the stricter the rules. The regulation defines the use of AI in the following areas as high risk, due to the potential impact on fundamental rights, safety and well-being:

• safety components in products subject to EU harmonisation legislation (or as stand-alone products) that are required to undergo a third-party conformity assessment under the same EU harmonisation legislation;
• biometrics, when used for remote identification, categorising individuals by sensitive attributes (such as race or religion) or emotion recognition, except where used to simply verify identity;
• critical infrastructure, when the AI is a safety component in areas such as digital infrastructure, traffic, water, gas, heating and electricity;
• education and vocational training, including access to education issues, evaluating learning outcomes, assessing education levels or monitoring behaviour during tests;
• employment, including recruiting, selecting candidates, making decisions on employment terms (promotions, terminations), task allocation or performance monitoring;
• essential services – AI systems used by public authorities for evaluating eligibility for public services (healthcare, benefits), credit scoring, insurance risk assessment and prioritising emergency responses;
• law enforcement – AI systems used for assessing crime risks, polygraphs, evaluating evidence reliability, predicting recidivism or profiling individuals for criminal investigations;
• migration and border control – AI systems used to assess risks related to migration, asylum and visa applications, or to detect and identify individuals in migration contexts;
• administration of justice and democratic processes – AI systems used by judicial authorities for legal research and interpretation or systems that could influence election outcomes.

The regulation prohibits the following AI practices with an unacceptable level of risk.

• Subliminal or deceptive techniques to manipulate individual or group behaviour, impairing their ability to make informed decisions and potentially causing harm.
• Exploiting vulnerabilities based on age, disability or socioeconomic situations to manipulate individuals or groups, leading to potential harm.
• Generating or manipulating non-consensual intimate material, such as realistic images, videos, audio or similar material showing an identifiable person’s intimate parts or that person engaged in sexually explicit activities, without that person’s explicit consent.
• Generating or manipulating child sexual abuse material.
• Social scoring, evaluating or classifying people based on behaviour or characteristics, resulting in unfair treatment unrelated to the context in which the data were collected or in a manner disproportionate to the behaviour’s severity.
• Criminal risk assessment, predicting the likelihood of committing a crime solely based on profiling or personality traits, except in objective, fact-based criminal investigations.
• Facial recognition database scraping from the internet or security cameras without specific targeting.
• Inferring emotions in sensitive areas, such as workplaces or educational institutions, unless used for medical or safety purposes.
• Biometric categorisation based on data to infer sensitive attributes like race, religion or political opinions, except for lawful use in law enforcement.
• Real-time biometric identification in public by law enforcement, unless strictly necessary for particular situations (e.g. finding missing persons, preventing imminent threats or identifying suspects of serious crimes). This must follow strict legal procedures, including prior authorisation, a limited scope and safeguards to protect rights and freedoms.

The regulation introduces disclosure obligations where a risk could arise from a lack of transparency around the use of AI:

• AI designed to impersonate humans (e.g. a chatbot) needs to inform the human it is interacting with;
• the output of generative AI needs to be marked as AI generated in a machine-readable way;
• in certain cases, the output of generative AI needs to be visibly labelled, namely deepfakes and text that is intended to inform the public of matters of public interest.

All other AI systems are deemed to have limited risk and therefore the regulation does not introduce any further rules.

Trustworthy use of large AI models

• General-purpose AI models are AI models that are trained on large amounts of data and can perform a wide range of tasks. They can be components of AI systems.
• The regulation introduces transparency obligations for providers of such general-purpose AI models, namely technical documentation, the provision of information to downstream developers of AI systems and the disclosure of data used in training the model.
• The most powerful general-purpose AI models can pose systemic risks. If a model meets a certain threshold of capability, the provider of that model must comply with additional risk management and cybersecurity obligations.

Governance

• The regulation sets up several governing bodies:
  • national competent authorities that oversee and enforce the rules for AI systems;
  • an AI office within the European Commission that coordinates the coherent application of the common rules across the EU and acts as a supervisor of general-purpose AI models; in certain cases, it also supervises and enforces the rules for AI systems based on such models and for AI systems that constitute or are integrated into very large online platforms or search engines.
• The EU Member States and the AI office cooperate closely via an AI board, comprising Member States’ representatives, to ensure the consistent and effective application of the regulation.
• The regulation sets up two advisory bodies for the AI office and the AI board:
  • a scientific panel of independent experts to provide scientific advice;
  • an advisory forum for stakeholders to provide technical expertise to the AI board and the Commission.
• The Commission has adopted the following implementing acts under the regulation:
  • Implementing Regulation (EU) 2025/454 sets out rules on the establishment and operation of the scientific panel of independent experts;
  • Implementing Regulation (EU) 2026/1755 lays down detailed arrangements for evaluations of general-purpose AI models and procedural safeguards for certain Commission proceedings.

Penalties

The fines for infringements are set as a percentage of the offending company’s annual turnover or a predetermined amount, whichever is higher. The fines for SMEs, including start-ups, and SMCs are subject to specific adjustments to ensure proportionality.

Transparency and protecting fundamental rights

Increased transparency applies to the development and use of high-risk AI systems:

• before a high-risk AI system is deployed by entities providing public services, its fundamental rights impact must be assessed;
• high-risk AI systems and entities using them must be registered in an EU database.

Innovation

The regulation provides for an innovation-friendly legal framework and aims to promote evidence-based regulatory learning. It envisages AI regulatory sandboxes, enabling a controlled environment in which innovative AI systems can be developed, tested and validated, including in real-world conditions. Furthermore, the regulation allows real-world testing of high-risk AI systems under certain conditions.

Evaluation and review

The Commission assesses the need for amendments to the list of high-risk uses of AI and the list of prohibited practices every year. By 2 August 2028, and every four years thereafter, the Commission will evaluate and report on the following:

• adding to or extending the list of high-risk categories;
• amendments to the list of AI systems requiring additional transparency measures;
• amendments to improve supervision and governance.