What Is the EU AI Act? Rules, Risk Levels and Key Dates Explained

The EU AI Act is the world's first comprehensive law governing artificial intelligence. It is Regulation (EU) 2024/1689, a European Union law that sorts AI systems by risk, bans the most dangerous uses, sets strict rules for high-risk systems and for the powerful general-purpose models behind tools such as ChatGPT, and requires transparency whenever people are dealing with AI. It entered into force on 1 August 2024 and is being switched on in stages, with the largest set of obligations, for high-risk AI, now due from 2 December 2027 after a delay agreed in 2026.
What is the EU AI Act in simple terms?
The AI Act is a single, horizontal law that covers almost every use of artificial intelligence in the European Union, rather than regulating one sector at a time. The European Commission describes it as the first-ever comprehensive legal framework on AI worldwide, and says its aim is to foster trustworthy AI in Europe. It sits alongside a wider package of measures, including the AI Continent Action Plan, the AI Innovation Package and the launch of AI Factories, which are intended to support the development of AI while protecting safety and fundamental rights.
At its heart is a risk-based approach. Instead of treating all AI the same, the Act defines four levels of risk and applies rules that match the danger. Most everyday AI faces no new rules at all. Systems that can affect people's health, safety, livelihoods or rights face the heaviest duties, and a small group of practices is banned outright.
The law was approved by the European Parliament and the Council in June 2024, published in the Official Journal on 12 July 2024 and entered into force on 1 August 2024. From the start it was designed to apply in stages over several years, so that businesses and regulators could prepare. One of those stages has since been moved back, which is why dates you may have seen in older guides no longer apply. The current timetable is set out below.
The four risk levels explained
The Act defines four levels of risk for AI systems. The level decides what, if anything, a provider or user of the system must do.
Unacceptable risk covers AI systems considered a clear threat to the safety, livelihoods and rights of people. These are banned. The Act prohibits nine practices, which are listed in the next section.
High risk covers AI use cases that can pose serious risks to health, safety or fundamental rights. These systems may still be sold and used, but only if they meet strict obligations on risk management, data quality, record-keeping, documentation, human oversight, accuracy and cybersecurity, and pass a conformity assessment before they go on the market.
Transparency risk covers systems where the danger is that people do not realise they are dealing with AI. The Act imposes specific disclosure duties, for example telling a user they are talking to a chatbot, and making AI-generated content identifiable.
Minimal or no risk covers the vast majority of AI systems currently in use, according to the Commission. The Act introduces no rules for these. The Commission gives AI-enabled video games and spam filters as examples.
What AI is banned under the AI Act?
The Act prohibits nine practices. The first eight became effective on 2 February 2025, and the ninth takes effect in December 2026, having been added by the 2026 AI Omnibus. The banned practices are:
- harmful AI-based manipulation and deception;
- harmful AI-based exploitation of vulnerabilities, for example linked to age or disability;
- social scoring by public authorities;
- individual criminal offence risk assessment or prediction, in other words AI that predicts who will commit a crime;
- untargeted scraping of the internet or CCTV material to create or expand facial recognition databases;
- emotion recognition in workplaces and education institutions;
- biometric categorisation used to deduce protected characteristics such as race, political opinions or sexual orientation;
- real-time remote biometric identification for law enforcement in publicly accessible spaces, with only narrow exceptions;
- AI systems that generate non-consensual sexually explicit and intimate content or child sexual abuse material, such as AI nudification apps.
The Commission has published guidelines on the prohibited practices and on the definition of an AI system to help organisations work out whether they are in scope. There is no compliance route for a banned practice: a system that falls within the list cannot be fixed with documentation or a conformity assessment.
What counts as high-risk AI?
High-risk AI falls into two groups. The first is AI used in the sensitive areas listed in Annex III of the Act: biometric identification and categorisation, critical infrastructure, education and vocational training, employment and worker management, access to essential private and public services such as credit scoring, law enforcement, migration, asylum and border control, and the administration of justice and democratic processes. The second is AI that acts as a safety component of products covered by existing EU product-safety law, listed in Annex I, for example AI in robot-assisted surgery or in machinery.
Real examples help to show the scale. A CV-sorting tool used in recruitment, a system that scores school exams, software that decides whether someone gets a loan, and a tool that automatically examines visa applications are all high-risk under the Act. A spam filter or a video game is not.
Before a high-risk system can be placed on the market, its provider must put in place a risk assessment and mitigation system, use high-quality datasets designed to minimise discriminatory outcomes, keep logs so results can be traced, produce detailed technical documentation, give the deployer clear information, and build in human oversight and a high level of robustness, cybersecurity and accuracy. Most high-risk systems also need a conformity assessment, CE marking, registration in the EU database and post-market monitoring once they are in use. Deployers, the organisations that actually use the systems, must ensure human oversight and monitoring, and providers and deployers must report serious incidents.
What are the rules for general-purpose AI models?
General-purpose AI, or GPAI, models are the foundation models that can perform a wide range of tasks and now sit underneath many AI systems. The Act has a separate regime for them, set out in Chapter V, which has applied since 2 August 2025.
All GPAI providers must produce technical documentation, adopt a policy to respect EU copyright law, and publish a summary of the content used to train the model. Models that may pose systemic risk face additional duties, including model evaluation, adversarial testing, serious-incident reporting and cybersecurity measures. The threshold for systemic risk is currently set by computing power used in training, at 10 to the power of 25 floating point operations.
To support compliance, the Commission published three instruments in July 2025: guidelines on the scope of the GPAI obligations, a voluntary GPAI Code of Practice drawn up by independent experts, and a template for the public summary of training content. The template asks providers to give an overview of the data used to train their models, including the sources it came from.
What must companies do about transparency?
The transparency rules took effect in August 2026. They require that people are told when they are interacting with an AI system, so a chatbot must make clear that it is a machine. Providers of generative AI must ensure AI-generated content is identifiable, and certain AI-generated content must be clearly and visibly labelled, namely deepfakes and text published to inform the public on matters of public interest.
The Commission has also published a voluntary Code of Practice on Marking and Labelling of AI-generated Content, with a set of icons that creators, publishers and other deployers can use to disclose the artificial nature of images, audio, including deepfakes, and text, plus guidelines on the transparency obligations themselves. For generative systems already on the market before 2 August 2026, a grace period for machine-readable marking runs to 2 December 2026.
Key dates and deadlines
The timetable has changed since the Act was first passed, so the dates below reflect the position as it stands in November 2026:
- 1 August 2024: the AI Act enters into force.
- 2 February 2025: the first eight prohibited practices and the AI literacy obligation take effect.
- 2 August 2025: the governance rules and the obligations for GPAI model providers take effect.
- 2 August 2026: the Act becomes generally applicable; the AI Office's enforcement powers and the Article 50 transparency duties take effect.
- 27 July 2026: the Digital Omnibus on AI enters into force and moves the high-risk dates back.
- 2 December 2026: the ninth prohibition, on AI that generates non-consensual intimate imagery and child sexual abuse material, takes effect, and the grace period for machine-readable marking of older generative systems ends.
- 2 December 2027: strict obligations for high-risk systems in the Annex III sensitive areas take effect.
- 2 August 2028: obligations for high-risk AI embedded in regulated products under Annex I take effect.
What changed in 2026? The AI Omnibus delay
The biggest change to the Act since it was passed is the Digital Omnibus on AI, part of the Commission's Digital Package on Simplification. The proposal was adopted on 19 November 2025, a political agreement was reached on 7 May 2026, and it entered into force on 27 July 2026 as Regulation (EU) 2026/1744.
The Omnibus moved the high-risk deadlines back. Obligations for standalone high-risk systems in the Annex III areas moved from 2 August 2026 to 2 December 2027, a delay of 16 months, and obligations for high-risk AI embedded in regulated products moved from 2 August 2027 to 2 August 2028. The Commission says the extra time is meant to ensure the rules apply when companies have the right support tools, such as standards, in place.
Not everything moved. The GPAI obligations, which have applied since August 2025, the transparency duties and the Commission's fining powers over GPAI providers all kept their original dates. The Omnibus also added the ninth prohibition on non-consensual intimate imagery, reinforced the AI Office's powers and centralised oversight of systems built on GPAI models, extended simplified requirements previously granted to small and medium-sized enterprises to small mid-cap companies, opened regulatory sandboxes more widely, including an EU-level sandbox, and clarified how the AI Act interacts with EU product safety law, in particular the Machinery Regulation.
What are the fines for breaking the EU AI Act?
The penalties are set by Article 99 of the Act and are tiered by severity. The most serious breaches, violating the Article 5 prohibitions, can be fined up to 35 million euros or 7% of a company's total worldwide annual turnover, whichever is higher. Breaches of most other obligations, including the high-risk system requirements, carry fines of up to 15 million euros or 3% of worldwide turnover. Supplying incorrect or misleading information to the authorities can be fined up to 7.5 million euros or 1% of turnover. For small and medium-sized enterprises, including start-ups, the fine is the lower of the fixed amount or the percentage, rather than the higher.
Enforcement is shared. From 2 August 2026 the AI Office and the authorities of the member states are responsible for implementing, supervising and enforcing the Act. The AI Office holds enforcement powers over GPAI models: it can request technical documentation, evaluate models, require corrective measures and issue fines. National market-surveillance authorities enforce the rules for most other AI systems. The AI Board, a Scientific Panel and an Advisory Forum steer and advise the Act's governance.
Does the EU AI Act apply in the UK?
The Act is EU law, so it does not apply in the UK as domestic law. But it has extraterritorial reach, in a similar way to the GDPR. A company based outside the EU is still in scope if it places an AI system or a GPAI model on the EU market, or if the output of its AI system is used in the EU. Non-EU providers of high-risk systems must also appoint an authorised representative established in the EU.
That matters for British businesses. A UK company selling an AI recruitment tool to a German employer, or a UK software firm whose AI service produces outputs used by people in the EU, can be caught by the Act even though it has no EU office. The UK has not passed an equivalent comprehensive AI law, so a UK firm operating on both sides of the Channel may face the EU's rules for its European customers and a different, sector-led regime at home.
What happens next?
The next milestone is 2 December 2026, when the ninth prohibition takes effect and the grace period for machine-readable marking ends. After that, attention turns to the run-up to December 2027, when the high-risk obligations land. The Commission is continuing to publish guidance, and the AI Act Service Desk provides information and support for implementation. A July 2026 action plan on cybersecurity and AI also sets out a coordinated approach to the cybersecurity challenges posed by the most advanced AI models, with the Commission planning a call to increase EU evaluation capacity for AI models before they are placed on the market, expected to be operational by 2027.
For organisations, the practical work is unchanged by the delay: map which AI systems are in use, check whether any fall into the Annex III high-risk areas or are GPAI models, and build the documentation, logging and human oversight the Act will require. The extra 16 months are a window to prepare, not a reason to shelve the work.
Sources: European Commission, AI Act; European Commission AI Act Service Desk, ai-act-service-desk.ec.europa.eu. Regulation (EU) 2024/1689 and Regulation (EU) 2026/1744 (the Digital Omnibus on AI).
Frequently asked questions
When did the EU AI Act come into force? The AI Act, Regulation (EU) 2024/1689, entered into force on 1 August 2024. Its rules apply in stages, with the general application date of 2 August 2026 and high-risk obligations now due from 2 December 2027.
Is ChatGPT covered by the EU AI Act? Yes. ChatGPT is built on a general-purpose AI model, so its provider faces the GPAI obligations that have applied since 2 August 2025, including technical documentation, a copyright policy and a public summary of training content, plus extra duties if the model is classed as posing systemic risk.
Does the EU AI Act apply to UK companies? It can. A UK company is in scope if it places an AI system on the EU market or if the output of its AI is used in the EU, and non-EU providers of high-risk systems must appoint an authorised representative in the EU.
What is the AI Omnibus? The Digital Omnibus on AI, Regulation (EU) 2026/1744, is the 2026 law that simplified the AI Act. It entered into force on 27 July 2026 and moved the high-risk obligations to 2 December 2027 for Annex III systems and 2 August 2028 for AI embedded in regulated products, and added a ninth banned practice.