The EU AI Act explained

The EU's Artificial Intelligence Act, Regulation (EU) 2024/1689, is the first broad law written specifically for AI. It does not license AI or ban it in general. It sets duties that depend on two questions: what the AI is used for, and what role your organisation plays in bringing it to people in the EU.

That design makes the Act narrower than its reputation for most businesses, and much heavier for a few. This explainer sets out the structure as it stands in October 2026, after the first set of amendments, using the Official Journal text and the European Commission's own guidance. It is legal context, not legal advice. National enforcement rules, sector laws and the facts of your system all matter, and anyone with a real compliance decision to make should take advice from a qualified lawyer.

What the Act covers, and who it reaches

The Act regulates two things. The first is an AI system, defined as a machine-based system that operates with some autonomy and "infers, from the input it receives, how to generate outputs" such as predictions, content, recommendations or decisions.[1] The second is a general-purpose AI model: a model that "displays significant generality" and can perform a wide range of distinct tasks, whatever way it is released.[1] A large language model is a model; the chat assistant or agent built on it is a system. That distinction runs through the whole law.

The Act's reach is not limited to EU companies. Article 2 applies it to providers who place AI systems or general-purpose models on the EU market "irrespective of whether those providers are established or located within the Union or in a third country", to deployers located in the EU, and to providers and deployers outside it in one further case:[1]

“where the output produced by the AI system is used in the Union”

— Regulation (EU) 2024/1689, Article 2(1)(c).[1]

There are carve-outs. The Act does not apply to AI used exclusively for military, defence or national security purposes, to AI developed solely for scientific research, to research and testing before a system is released, or to people using AI "in the course of a purely personal non-professional activity".[1] Systems released under free and open-source licences are largely outside it too, unless they are high-risk, banned, or fall under the transparency rules.[1] Our guide to open versus closed AI models covers how openness changes the duties for model providers.

Four tiers for systems, one track for models

The Commission describes the Act as defining four levels of risk for AI systems.[2] General-purpose models sit beside that pyramid rather than inside it, because their duties attach to the model regardless of how anyone later uses it.[3]

AI Chronicle

Prohibited practices

Article 5 bans a short list of practices outright. In summary, it bans AI that uses subliminal, manipulative or deceptive techniques to distort behaviour in ways that cause significant harm; AI that exploits vulnerabilities due to age, disability or social or economic situation; social scoring that leads to unjustified or out-of-context detrimental treatment; predicting an individual's risk of committing a crime based solely on profiling or personality traits; building facial recognition databases by untargeted scraping of the internet or CCTV; inferring emotions in workplaces and education institutions, except for medical or safety reasons; biometric categorisation to infer race, political opinions, trade union membership, religious or philosophical beliefs, sex life or sexual orientation; and real-time remote biometric identification in public spaces for law enforcement, outside narrow exceptions.[1]

These eight have applied since 2 February 2025.[2] The Digital Omnibus added a ninth: AI systems that generate realistic intimate or sexually explicit images, video or audio of an identifiable person without their explicit consent, and AI that generates child sexual abuse material. That ban applies from 2 December 2026.[4][5]

High-risk systems

High-risk status comes from use, not from technical sophistication. There are two routes. An AI system is high-risk if it is a safety component of, or itself is, a product covered by the EU product laws listed in Annex I, such as machinery or medical devices, and that product needs third-party conformity assessment. It is also high-risk if it is intended for one of the uses listed in Annex III.[1]

Annex III covers eight areas: biometrics; critical infrastructure; education and vocational training; employment and worker management; access to essential private and public services, including creditworthiness checks and life and health insurance pricing; law enforcement; migration, asylum and border control; and the administration of justice and democratic processes.[1] In employment, for example, it lists AI intended "to analyse and filter job applications, and to evaluate candidates".[1] The Act regulates how such tools are used, not what AI does to employment overall; for that, see our explainer on AI and jobs: what the evidence says.

There is an exception. Under Article 6(3), an Annex III system is not high-risk if it does not pose a significant risk of harm, for example because it performs only a narrow procedural task or a preparatory task. The exception never applies where the system profiles people, and a provider relying on it must document its assessment and register the system.[1] The Commission published draft guidelines on classification, with practical examples, for consultation until 23 July 2026; they had not been formally adopted when we checked.[6]

Transparency duties

Article 50 sets disclosure duties that apply whether or not a system is high-risk. Providers must design systems that interact directly with people, such as chatbots, so people are told they are dealing with AI, unless that is obvious. Providers of generative systems must mark outputs "in a machine-readable format" so they are detectable as AI-generated. Deployers must disclose deepfakes, and AI-generated text published to inform the public on matters of public interest, unless that text has had human review or editorial control and someone holds editorial responsibility.[1] The Commission's FAQ names "chatbots, AI agents, and avatars" as examples of systems that interact with people.[7]

Article 50 has applied since 2 August 2026. The only grace period, to 2 December 2026, covers the machine-readable marking duty for generative systems already on the market before 2 August 2026.[4][7] Deployers' disclosure duties were not deferred. Our guide to AI for marketing works through what the labelling rules mean for campaigns, and our explainer on content credentials covers one way provenance can be recorded. Marking is not the same as detection: our comparison of AI detectors vs. content credentials explains why a detector's guess cannot stand in for a machine-readable mark.

Minimal or no risk

Everything else falls into the base of the pyramid. In the Commission's words, "The AI Act does not introduce rules for AI that is deemed minimal or no risk", and it says the vast majority of AI systems used in the EU fall there, giving spam filters and AI-enabled video games as examples.[2] Other laws, such as data protection and consumer law, still apply.

General-purpose AI models

Providers of general-purpose AI models must keep technical documentation for the AI Office and national authorities, give downstream system providers the information they need, put in place a policy to comply with EU copyright law, including rights reservations under the text and data mining exception, and publish a summary of the content used for training, using a Commission template.[1][2] Providers of openly released models with public weights are exempt from the two documentation duties, but not from the copyright and training-summary duties.[1]

A model is presumed to have systemic risk when the compute used to train it exceeds 10^25 floating-point operations, and the Commission can designate others based on criteria such as reach and capabilities.[1][8] Providers of these models must also evaluate them, including adversarial testing, assess and mitigate systemic risks, report serious incidents, and ensure adequate cybersecurity.[1] Copyright questions about training data are bigger than the Act's disclosure duties; we cover them in our guide to AI training and copyright.

Provider or deployer: the roles that decide your duties

Most obligations attach to a role rather than to a company type. The definitions are short.

Source: Regulation (EU) 2024/1689, Article 3(3)–(7).[1] The examples follow the Commission's own illustration of a provider and a deployer.[8]

Roles can change. Under Article 25, a deployer, distributor, importer or other third party becomes the provider of a high-risk system if it puts its own name on one, substantially modifies one, or changes the intended purpose of a system, "including a general-purpose AI system", so that it becomes high-risk. The original provider then stops being the provider of that system but must cooperate, and the Omnibus spelled out that this includes sharing technical documentation, known limitations and failure modes, and targeted technical access for testing.[1][4] A business that turns a general-purpose assistant into a candidate-ranking tool should assume it has become a provider.

The same logic applies to models. The AI Office's guidance says companies that fine-tune or modify a general-purpose model "may become providers of new models", and gives an indicative threshold: the modification uses more than a third of the original model's training compute.[3] Below that, a company building on a model is a provider or deployer of an AI system, not of the model. Our analysis of who is accountable when an AI agent acts traces how these roles map onto a single agent action.

What each role must do

The heaviest duties belong to providers of high-risk systems. Article 16 requires them to meet the Act's requirements, which cover risk management, data quality and governance, technical documentation, logging, information for deployers, human oversight, and accuracy, robustness and cybersecurity. They must also run a quality management system, keep documentation and logs, undergo conformity assessment before release, draw up an EU declaration of conformity, affix the CE marking, register in the EU database and take corrective action when needed.[1][2] The Omnibus lets small and medium-sized enterprises and small mid-cap companies use a simplified technical documentation form.[4]

Deployers of high-risk systems have a shorter list under Article 26. They must use the system according to its instructions, assign human oversight to people "who have the necessary competence, training and authority, as well as the necessary support", make sure input data under their control is relevant, monitor the system and report serious incidents, keep logs under their control for at least six months, inform workers before using the system at work, and tell people when a high-risk system makes or assists decisions about them.[1] Public bodies, private entities providing public services, and deployers using AI for credit scoring or life and health insurance pricing must also carry out a fundamental rights impact assessment.[1] People affected by certain decisions based on a high-risk system's output can ask the deployer for an explanation of its role.[1]

Source: Regulation (EU) 2024/1689, Articles 4, 5, 16, 26, 50, 53, 55, 111 and 113, as amended by Regulation (EU) 2026/1744.[1][4]

The AI literacy duty changed. The original text required providers and deployers to ensure, "to their best extent, a sufficient level of AI literacy" among staff.[1] The amended Article 4 requires them to "take measures to support the development of AI literacy", and adds that this "does not require providers or deployers to guarantee any specific level of AI literacy of any individual."[4]

The timeline as of October 2026

The Act entered into force on 1 August 2024 and applies in stages.[2] The Digital Omnibus on AI, Regulation (EU) 2026/1744, was proposed on 19 November 2025, published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026.[4][9] Its main effect was to delay the high-risk rules. The Commission explained why:

“The delayed availability of the standards puts in jeopardy the successful entry into application of the high-risk rules on 2 August 2026.”

— European Commission, Navigating the AI Act (Q&A, last updated 7 August 2026).[8]

The amended Article 113 now sets fixed dates for the high-risk chapter: 2 December 2027 for systems classified under Annex III, and 2 August 2028 for systems classified under Annex I.[4] The Commission's proposal had linked the start date to the availability of standards, with those dates as the outer limit; the adopted text simply fixes them.[4][10]

AI Chronicle

Four other transitional rules matter in practice:

• Existing high-risk systems. High-risk systems placed on the market before the high-risk chapter applies are covered only if their design changes significantly after that date. Systems intended for public authorities must comply by 2 August 2030 in any case.[4]
• Existing general-purpose models. Models placed on the market before 2 August 2025 must comply by 2 August 2027.[1]
• Existing generative systems. Providers of systems generating synthetic content that were on the market before 2 August 2026 have until 2 December 2026 for Article 50(2) marking.[4]
• Enforcement. The Commission's timeline says enforcement at national and EU level started on 2 August 2026 for model rules, prohibitions, transparency rules and AI literacy.[5]

The General-Purpose AI Code of Practice

The Code of Practice is a voluntary tool, written by independent experts in a multi-stakeholder process, that shows model providers one way to meet their obligations.[11] It was published on 10 July 2025 and has three chapters. The Transparency and Copyright chapters are for all providers of general-purpose models and include a Model Documentation Form; the Safety and Security chapter is only for providers of models with systemic risk.[11] The Commission and the AI Board have assessed it as adequate.[11]

Signing it is not the same as being certified. The AI Office's Q&A explains that adherence to the Code:

“becomes a means to demonstrate compliance with the AI Act, while not providing a presumption of conformity”

— AI Office, General-Purpose AI Models in the AI Act – Questions & Answers.[3]

Providers who do not sign must show that their alternative means of compliance are adequate, for instance through a gap analysis against the Code. The AI Office can take Code commitments into account when monitoring compliance and as a mitigating factor when setting fines.[3] As of the Commission's list updated on 7 October 2026, signatories included Amazon, Anthropic, Google, IBM, Microsoft, Mistral AI and OpenAI, among others; xAI had signed only the Safety and Security chapter.[11]

The Code does not cover products built on these models. A business that uses a model through an API is not a signatory candidate unless it is itself a model provider. A separate Code of Practice on transparency of AI-generated content supports Article 50 compliance for generative systems.[7]

Penalties and who enforces them

Member States set the penalties for AI systems within ceilings fixed by Article 99. For each category, an undertaking faces whichever is higher of a fixed amount and a share of total worldwide annual turnover for the preceding financial year:[1]

Source: Regulation (EU) 2024/1689, Articles 99 and 101.[1]

For SMEs and start-ups, each ceiling is whichever is lower, and the Omnibus extended the lower-of rule to small mid-cap companies for the second and third categories.[1][4] Authorities must consider the nature, gravity and duration of the infringement, the size of the operator and other circumstances.[1]

Enforcement is split. National market surveillance authorities oversee AI systems; the AI Office, part of the Commission, enforces the rules for general-purpose models and can request documentation, evaluate models, require corrective measures and issue fines.[2][12] The Omnibus extended the AI Office's oversight to certain systems built on general-purpose models and to AI embedded in very large online platforms and search engines.[9]

What a business outside the EU needs to do

Being based outside the EU does not take you out of scope. The Act applies to providers placing systems or models on the EU market wherever they are established, and to providers and deployers abroad whose systems' output is used in the EU.[1] A practical sequence for a non-EU business:

1. Inventory your AI. List the systems you build and the systems you use, including features built on third-party models, and note which touch people in the EU or produce outputs used there.
2. Find your role for each. Are you putting a system out under your own name (provider), using one professionally (deployer), or both? Check whether rebranding, substantial modification or a new purpose has made you a provider under Article 25.[1]
3. Screen against the bans. Make sure no product or internal use falls under Article 5, including the new ban from 2 December 2026.[1][4]
4. Check Article 50 now. Chatbots must disclose that users are talking to AI; generators must mark outputs (by 2 December 2026 for systems launched before 2 August 2026); deepfakes and public-interest text need labels.[1][7]
5. Classify against Annex III and Annex I. If something may be high-risk, plan for 2 December 2027 or 2 August 2028. Document any Article 6(3) assessment that it is not.[1][4]
6. Appoint an EU representative where required. Non-EU providers of high-risk systems must appoint an authorised representative established in the EU, by written mandate, before making the system available in the EU. Non-EU providers of general-purpose models must do the same before placing a model on the EU market, unless the open-source exemption applies and the model has no systemic risk.[1]
7. Set up AI literacy measures and records. Train the staff who operate AI, and keep the documentation the Act asks for, so you can answer an authority's request.

A worked example: one company, five AI uses

This is a hypothetical example. It describes no real company, product or regulatory decision. The classifications are illustrative readings of the text, not legal conclusions, and a real assessment would depend on facts this sketch leaves out.

A 60-person software company based in Canada sells business software to customers in several countries, including EU member states. It has no EU office. It uses one vendor's general-purpose model through an API.

Two points in the table deserve a closer look. The CV-ranking add-on scores applicants, which very likely involves profiling, so the Article 6(3) exception would not be available; that is our reading, and the Commission's final guidelines may refine it.[1][6] The EU employers who buy the add-on become deployers, with their own duties to assign human oversight, keep logs for at least six months, inform workers and tell candidates that a high-risk system is in use.[1] The model vendor, meanwhile, carries the general-purpose model duties, and under Article 53 owes the company the information it needs to understand the model's capabilities and limitations.[1]

What the company should not conclude is that the chatbot's "AI can make mistakes" footer covers its obligations. Disclosure that a user is talking to AI is a separate duty from accuracy, and accountability for what an agent does stays with the business either way.

Questions readers commonly ask

Does the AI Act apply if we only use AI tools internally?

Yes, as a deployer, but for most internal uses the duties are light. You must avoid banned practices and take measures to support staff AI literacy.[4] Heavier duties arise only if the use is high-risk, such as screening job applicants, or if you publish deepfakes or public-interest text without editorial review.[1]

Are the new high-risk dates final?

They are in force. The Digital Omnibus fixed 2 December 2027 for Annex III systems and 2 August 2028 for Annex I systems in the amended Article 113.[4] Any further change would need new legislation.

Do we have to sign the General-Purpose AI Code of Practice?

Only providers of general-purpose AI models can sign, and signing is voluntary. It is one way to demonstrate compliance, not a presumption of conformity.[3] Businesses that build on a model through an API are usually not model providers.

How does the AI Act relate to the GDPR?

The AI Act does not replace data protection law; Article 2(7) says it does not affect the GDPR.[1] Where a deployer must do both a data protection impact assessment and a fundamental rights impact assessment, the Omnibus lets it cross-reference the first in the second.[4]

Is an AI agent treated differently from a chatbot?

Not as a category. The Act regulates by use and role, so an agent is assessed like any other AI system. The Commission lists AI agents among systems that must disclose they are AI when interacting with people.[7] If an agent is used for an Annex III purpose, the high-risk rules apply.

The next step

The Act's model duties include a copyright policy and a public training-data summary, but they leave the underlying legal questions open. Next, read AI training and copyright: the important distinctions.

About these sources

All pages were accessed on 16 October 2026. Legal texts are cited for what they state, not as legal advice. Where the Commission's summaries and the Official Journal text differ, the Official Journal text was followed. The software company and its products in the worked example are hypothetical, and the classifications in it are illustrative.

Related reading

• Who is accountable when an AI agent acts?
• How to spot AI-generated content: signals and limits