The EU AI Act Has Applied in Stages Since 2 February 2025, and Its Rules for High-Risk Systems Now Wait Until 2 December 2027 and 2 August 2028

Most of the EU’s AI Act is already live. The bans on eight kinds of AI use have applied since 2 February 2025, the rules for general-purpose AI models since 2 August 2025, and the general application date of 2 August 2026 has passed, which started the transparency duties and the national and EU enforcement of the rules already in force. The rules for high-risk systems have not begun. The European Commission says they now apply from 2 December 2027 for the uses listed in Annex III and from 2 August 2028 for AI built into regulated products.

This is a plain-language summary of Regulation (EU) 2024/1689, not legal advice. Article numbers and dates come from the Commission’s AI Act Service Desk, whose text is the consolidated version as at 27 July 2026, and from Commission pages last updated between 28 April and 7 October 2026.

The dates, in the order they arrive

The Regulation was signed on 13 June 2024 and published in the Official Journal on 12 July 2024, according to the European Parliament’s Legislative Train page. Article 113 sets entry into force for the twentieth day after publication, and the Commission’s AI Act page gives the date as 1 August 2024. The Service Desk timeline then lists what applies and when:

• 2 February 2025: the definitions, AI literacy and the prohibited practices.
• 2 August 2025: obligations for providers of general-purpose AI models. Member States also had to name national authorities and adopt penalty laws by this date.
• 2 August 2026: most of the remaining rules, including the Article 50 transparency duties. Enforcement starts at national and EU level for general-purpose models, prohibitions, transparency rules and AI literacy.
• 2 December 2026: a new ban on AI that generates non-consensual sexual deepfakes and child sexual abuse material, and a transitional deadline for some synthetic-content systems already on the market before 2 August 2026.
• 2 August 2027: each Member State should have at least one AI regulatory sandbox (a supervised testing space) running. Article 111(3) also gives general-purpose models placed on the market before 2 August 2025 until this date to comply.
• 2 December 2027: the high-risk rules for the uses in Annex III.
• 2 August 2028: the high-risk rules for AI in products covered by Annex I, such as lifts, toys and machinery.

What moved, and where it stands on 15 October 2026

The original text set 2 August 2026 for Annex III systems and 2 August 2027 for Annex I systems, as the law firm Gibson Dunn reported on 27 May 2026. The Commission proposed the Digital Omnibus on AI on 19 November 2025. Its “Navigating the AI Act” FAQ says the proposal linked the start of the high-risk rules to the availability of support measures such as standards.

A Commission news item dated 27 July 2026 says the AI Omnibus entered into force that day and gives the two new dates. The same item lists other changes: simplified obligations that applied to small firms now also cover small mid-cap companies, a new EU-level regulatory sandbox, a simpler AI literacy requirement, permission to process special categories of personal data to detect and correct bias, and wider AI Office oversight of some systems built on general-purpose models and of AI embedded in very large online platforms and search engines.

The newest pages opened for this article are the Commission’s AI Act page (updated 3 August 2026) and its “Navigating the AI Act” FAQ (7 August 2026). Both carry the 2 December 2027 and 2 August 2028 dates and describe the Omnibus as in force. None of the Commission pages opened here, the latest updated on 7 October 2026, mentions a further postponement.

Four tiers of risk

The Commission’s AI Act page sorts AI systems into four levels. Unacceptable-risk uses are banned. High-risk systems face strict obligations before they reach the market. Transparency-risk systems, such as chatbots and deepfake generators, must disclose that AI is involved. Everything else is minimal risk and gets no extra rules; the page names video games and spam filters as examples. General-purpose AI models carry their own set of duties alongside these tiers.

The banned list

Article 5(1) prohibits, in points (a) to (h), the following:

• hidden or deceptive techniques that materially distort behaviour and cause significant harm
• exploiting a person’s age, disability or social or economic situation to the same effect
• social scoring that leads to unfair or disproportionate treatment
• judging a person’s risk of committing a crime from profiling or personality traits alone
• building facial recognition databases by scraping images from the internet or CCTV without a specific target
• inferring emotions at work or school, except for medical or safety reasons
• using biometric data to infer race, political opinions, union membership, religious beliefs, sex life or sexual orientation
• real-time biometric identification in public spaces for law enforcement, except in narrow cases such as searching for abduction victims, preventing imminent serious threats or locating suspects of serious offences

The Omnibus added two new points, (ba) and (bb), covering intimate or sexually explicit deepfakes and child sexual abuse material. The Commission page counts the new ban as the ninth prohibition. Under Article 113 it applies from 2 December 2026. Article 99(3) sets the top fine for a banned practice at 35 million euros or 7% of worldwide annual turnover, whichever is higher.

High-risk systems

Article 6 gives two routes to the high-risk label. One is an AI system that is a safety component of a product, or is itself a product, covered by the EU harmonisation laws listed in Annex I, where that product needs a third-party conformity assessment. The other is any system listed in Annex III.

Annex III names eight areas: biometrics, critical infrastructure, education and vocational training, employment and worker management, access to essential services and benefits (including credit scoring), law enforcement, migration, asylum and border control, and the administration of justice and democratic processes.

Article 6(3) lets a provider conclude that an Annex III system is not high-risk if it only does a narrow procedural task, improves the result of a human activity already completed, detects patterns without replacing human assessment, or does a preparatory task. A system that profiles people is always high-risk. A provider who takes the exemption route must document the assessment and register it.

A provider of a high-risk system must, under Article 16, run a quality management system, keep documentation and automatically generated logs, put the system through conformity assessment before sale or use, draw up an EU declaration of conformity, apply the CE marking, register it and take corrective action when needed. None of those duties bind until the dates above.

Transparency duties

Article 50 applies from 2 August 2026, according to the Commission. It has four main duties:

• Providers must design chatbots and similar systems so people know they are talking to an AI, unless that is obvious.
• Providers of systems that generate synthetic audio, image, video or text must mark the output in a machine-readable form so it can be detected as AI-generated.
• Deployers of emotion recognition or biometric categorisation systems must tell the people exposed to them.
• Deployers must disclose deepfakes, and AI-generated text published to inform the public on matters of public interest, unless the text has human editorial review and someone holds editorial responsibility.

Artistic, satirical and fictional works get a lighter form of disclosure. The Service Desk timeline gives some synthetic-content systems already on the market before 2 August 2026 until 2 December 2026 to meet the marking duty. How these labelling rules sit beside India’s and the US’s is set out in our piece on election deepfakes.

General-purpose AI models

Under Article 53, every provider of a general-purpose AI model must keep technical documentation for the AI Office and national authorities, give downstream providers information they need, keep a policy to comply with EU copyright law, and publish a sufficiently detailed summary of the content used for training, using a template from the AI Office. Models released under a free and open-source licence are exempt from the first two duties unless they carry systemic risk, and Article 53 names only those two, so the copyright policy and training summary still apply to them. Courts in India, Germany and the US have each ruled on training and copyright in 2026 without settling it, as this earlier article sets out.

A model is presumed to have systemic risk when the compute used to train it exceeds 10^25 floating-point operations, the Commission’s FAQ says. The AI Office’s Q&A says a provider must notify the Commission within two weeks of meeting that threshold. Providers of systemic-risk models must also evaluate the model, assess and mitigate systemic risks, report serious incidents and secure the model against cyberattack.

The Commission’s General-Purpose AI Code of Practice, delivered on 10 July 2025, is a voluntary way to show compliance, with chapters on transparency, copyright, and safety and security. Its page, updated on 7 October 2026, lists 22 companies as signatories, among them Amazon, Anthropic, Google, IBM, Microsoft, Mistral AI and OpenAI. xAI signed only the safety and security chapter, so the page says it must show compliance with the other duties by other means.

Under the AI Office’s Q&A, the Commission can fine a model provider up to 3% of worldwide turnover or 15 million euros, whichever is higher. The timeline puts the start of that enforcement at 2 August 2026.

Fines for everyone else, and companies outside the EU

Article 99 sets three fine ceilings for the Act’s other breaches: 35 million euros or 7% of turnover for banned practices, 15 million euros or 3% for breaches of listed obligations of operators and notified bodies, and 7.5 million euros or 1% for giving authorities incorrect or misleading information. For small and medium enterprises, each fine is capped at the lower of the two figures. The Omnibus added the same lower-of cap for small mid-cap companies.

Article 2 says the Regulation applies to providers who place AI systems or general-purpose models on the EU market whether they are established in the Union or in a third country. It also covers providers and deployers based outside the Union where the system’s output is used in the Union. The Commission’s FAQ describes the same reach for actors inside and outside the EU. None of the pages opened for this article names India or any other country, so an Indian company’s position turns on those tests: whether it places a system or model on the EU market, or whether its system’s output is used there.

The Commission’s timeline shows the next fixed dates as 2 December 2026, 2 August 2027, 2 December 2027 and 2 August 2028.