EU AI Act

The EU Artificial Intelligence Act is a regulation of the European Union that sets rules for the development, placing on the market, and use of AI systems. It was proposed in 2021, adopted in June 2024, and entered into force on 1 August 2024. It is the first comprehensive AI law adopted by a major jurisdiction 1. The Commission presents it as a risk-based framework that protects health, safety, and fundamental rights while supporting innovation 2.

The Act does not regulate algorithms in general. It focuses on particular uses, classified by risk, and on general-purpose models that can be adapted to many uses. Some practices are prohibited outright; others are permitted subject to obligations; most are left unregulated apart from transparency rules.

The law matters beyond Europe because it applies to providers wherever they are established if their systems are placed on the EU market, and because other jurisdictions have used it as a template. This entry covers its scope, obligations, timetable, enforcement, and the criticism it attracts from opposite directions.

The Commission proposed the Act in April 2021, after a white paper and a public consultation. The proposal combined product-safety law with fundamental-rights protection: AI systems would be treated like other products for safety purposes, with additional rules for uses that affect rights.

The negotiations took longer than expected. The European Parliament and the Council reached a political agreement in December 2023, after debates over biometric surveillance, foundation models, and the obligations of open-source developers. The final text was adopted in March 2024, signed in June, and published in the Official Journal in July 1.

The Act is a regulation rather than a directive, so it applies directly in member states without national transposition. That choice centralises the rules but leaves implementation to national authorities, which must designate market surveillance bodies and provide them with resources.

Implementation has been contested since adoption. Standards work has run behind schedule, and in 2026 the co-legislators agreed to postpone the high-risk obligations and to simplify several overlapping requirements 3 4.

The Act applies to providers, deployers, importers, and distributors of AI systems in the Union, and to providers established outside the Union where the system's output is used in the Union. It covers systems placed on the market, put into service, or used in the EU, regardless of where they were developed 1.

Its definition of an AI system is broad: a machine-based system that infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. The definition is deliberately technology-neutral, and it excludes systems that operate only on rules defined by natural persons.

There are significant exclusions. Systems developed or used exclusively for military, defence, or national security purposes are outside the Act; so is pure scientific research and development, and personal non-professional use. Open-source components are largely exempt unless they are placed on the market as high-risk systems or fall under prohibited practices.

The scope is therefore narrower than the phrase AI law suggests. Much of what governments do with AI in security and defence is untouched, and the treatment of research means that laboratories developing systems are not regulated until they deploy them 5.

The Act classifies practices into four tiers. At the top are prohibited practices, listed in Article 5: manipulation that causes significant harm, exploitation of vulnerability, social scoring by public authorities, predictive policing based on profiling, untargeted scraping of facial images, emotion recognition in workplaces and schools, and most real-time remote biometric identification in public spaces. The 2026 amendment added applications that create non-consensual sexual imagery 3.

Next are high-risk systems, defined by annexes rather than by a general test. Annex I covers AI used as a safety component in products regulated by EU safety law, such as machinery, medical devices, and vehicles. Annex III lists uses in biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, and the administration of justice. Providers can argue that a listed system does not pose a significant risk, but the exception does not apply where the system performs profiling 1.

Below that are systems subject to transparency obligations: chatbots must disclose that they are not human, synthetic content must be marked, and deepfakes must be labelled. Emotion recognition and biometric categorisation systems have their own disclosure duties.

The remaining systems are unregulated by the Act, though other law may apply. Obligations scale with assessed risk, and most AI is left alone 6.

Providers of high-risk systems carry most of the burden. They must operate a risk-management system across the lifecycle, apply data governance and testing, draw up technical documentation, keep logs, provide instructions for use, design for human oversight, and meet standards for accuracy, robustness, and cybersecurity. They must also run a quality-management system, undergo conformity assessment, affix CE marking, register the system in the EU database, and monitor it after release 1.

Deployers have duties that depend on context. They must use the system according to its instructions, assign competent human oversight, monitor for risks, and keep logs. Public bodies and certain private operators must carry out a fundamental-rights impact assessment before using a high-risk system, and employers must inform workers and their representatives before deploying one in the workplace.

Importers and distributors must verify that the provider has complied, and they become subject to provider obligations if they place a system on the market under their own name. The value chain is covered deliberately: a distributor cannot avoid responsibility by not being the developer.

These obligations are familiar from EU product-safety law, and that is the point. The Act extends the machinery of CE marking, technical files, and market surveillance to software, which is a novel application of an old regulatory model 1.

General-purpose AI models are regulated separately, in a chapter added during the negotiations to address systems that are trained on broad data and can perform many tasks. All providers must maintain technical documentation, supply information to downstream providers, publish a policy on copyright compliance, and provide a summary of training content 1.

Models with systemic risk carry additional obligations: evaluation and adversarial testing, incident reporting, cybersecurity measures, and disclosure of energy consumption. A model is presumed to have systemic risk if the cumulative compute used in training exceeds 10^25 floating point operations, a threshold the Commission can revise.

The obligations apply to providers wherever they are established, and the Commission's AI Office oversees them, supported by a scientific panel and an advisory forum 7. A code of practice, published in July 2025 and drawn up with industry and civil society, sets out ways to demonstrate compliance; adherence is voluntary but gives providers a presumption of conformity and reduces their administrative burden 8.

Open-weight models are not exempt. The Act allows providers to be exempted from some documentation duties where the model is released under a free and open-source licence, but the systemic-risk obligations and the copyright policy requirement remain, a provision open-source advocates criticised.

The Act applies in stages. Prohibitions and the obligation to ensure AI literacy took effect on 2 February 2025. Rules for general-purpose models, governance, and penalties applied from 2 August 2025. The high-risk regime was originally scheduled for 2 August 2026, with product-embedded systems following in 2027.

In 2026 the co-legislators agreed to postpone the high-risk obligations. Stand-alone high-risk systems listed in Annex III are now to be regulated from 2 December 2027, and systems embedded in products covered by sectoral safety law from 2 August 2028 3 4. Watermarking obligations were moved to 2 December 2026, and the ban on nudifier applications applies from the same date.

The postponement followed complaints that the standards needed to demonstrate compliance would not be ready, and that the original timetable would create legal uncertainty. The Commission had proposed a delay; the Parliament and Council negotiated the final dates and added the nudifier ban.

The timetable is now a political variable. Further delays are possible, and the precedent matters: a jurisdiction that postpones its own deadlines signals that enforcement is negotiable.

Enforcement is split. National market surveillance authorities supervise AI systems in their territory, conduct checks, and can order withdrawal or recall. The Commission's AI Office enforces the rules for general-purpose models and coordinates cross-border cases. A European AI Board brings together national authorities, and notified bodies assess conformity for high-risk systems 7.

Penalties are set in Article 99. The highest tier, up to 35 million euros or 7 per cent of worldwide annual turnover, applies to prohibited practices. Other breaches of obligations carry up to 15 million euros or 3 per cent, and supplying incorrect information carries up to 7.5 million euros or 1 per cent. For companies, the percentage is usually the binding constraint; for smaller firms, the fixed amount may be.

Individuals do not have a direct right to sue under the Act. Complaints go to market surveillance authorities, and remedies for harm depend on other law, including product liability and data protection. That design is one of the features critics point to when they argue that enforcement is administrative rather than rights-based.

Resources are the practical constraint: market surveillance authorities are small, the expertise required is specialised, and the number of systems is large. The Act gives authorities powers; whether they can use them at scale is another question 5.

Industry criticism focuses on cost, uncertainty, and timing. High-risk compliance requires documentation, testing, quality systems, and third-party assessment, which small companies say they cannot afford. The Act provides some support for SMEs, but the fixed costs of conformity assessment do not scale down easily.

The delay in harmonised standards has been the sharpest practical complaint. Without published standards, providers cannot be sure what conformity looks like, and notified bodies cannot assess against a stable benchmark. The postponement of the high-risk rules was, in effect, an admission that the compliance infrastructure was not ready 3.

A second complaint is legal uncertainty. Terms such as significant risk, human oversight, and sufficiently accurate are defined in the Act but require interpretation, and the guidance that would settle them has arrived slowly. Companies face a choice between over-compliance, which is expensive, and under-compliance, which is risky.

A third is overlap. A single product can be subject to the AI Act, the Medical Device Regulation, the General Data Protection Regulation, and sectoral rules. The 2026 amendment tried to reduce duplication for machinery and to narrow the definition of a safety component, but the general problem of layered obligations remains 3.

Civil-society criticism is that the Act is too weak where it matters. The prohibitions on biometric identification contain exceptions for law enforcement, and the list of prohibited practices was narrowed during negotiation. Remote biometric identification in public spaces remains available under judicial or administrative authorisation for a set of serious crimes.

High-risk classification depends on annexes that were lobbied over, and many systems are assessed by the provider itself rather than by an independent body. The fundamental-rights impact assessment applies only to certain deployers, and its results are not required to be published in full.

Access to justice is limited. Because the Act does not create individual rights or a direct complaint mechanism, people affected by a system depend on regulators to act. Critics argue that this makes enforcement remote and that affected communities have little influence over whether a case is pursued 5.

The general-purpose model regime is also seen as light. Transparency about training data is summary-level, the systemic-risk threshold is a compute proxy that may miss capability, and the code of practice is voluntary. For models that underpin the most widely used systems, the Act relies on documentation and disclosure rather than substantive limits.

The first disagreement is about whether the Act is a model or a cautionary tale. Supporters describe it as a credible attempt to regulate a fast-moving technology without banning it, and point to its influence on legislation elsewhere. Critics describe a law that was weakened during negotiation, delayed during implementation, and may end up regulating paperwork rather than behaviour 5.

The second is about competitiveness. Industry groups argue that compliance costs push AI development outside Europe. Others note that Europe's problem is investment and compute, not regulation, and that a predictable legal framework is an advantage for firms selling into the single market.

The third is about the risk-based approach itself. A categorical list is predictable for companies but rigid in the face of new uses; a general risk standard is flexible but uncertain. The Act tries to combine both with annexes and a derogation, and the result satisfies neither camp completely.

The fourth is about enforcement. The Brussels effect argument holds that EU rules become global standards because companies build one compliant product. The opposite argument holds that enforcement will be slow and selective, that national authorities have different priorities, and that the Act's influence will be rhetorical. The next few years of decisions, fines, and standards will settle more than any commentary can.