EU AI Act
The EU AI Act regulates AI systems by risk level, with prohibited practices, high-risk obligations, transparency rules and duties for general-purpose AI models.
Overview
The EU Artificial Intelligence Act, Regulation (EU) 2024/1689, is the European Union's law on artificial intelligence. It regulates AI systems according to the risk they pose. A short list of practices is banned outright. AI used in sensitive areas, such as hiring, credit, education, critical infrastructure and certain public services, is classed as high-risk and must meet strict requirements. Some systems, such as chatbots and generators of synthetic content, carry transparency duties. Providers of general-purpose AI models, the large models that many applications are built on, have their own obligations. Most other AI systems face no new specific rules.
An everyday analogy is how a country regulates vehicles. Some things are simply forbidden on public roads. Lorries and buses need special licences, inspections and logbooks because a failure can hurt many people. Ordinary cars need lights and number plates so others know what they are dealing with. Bicycles need very little. The AI Act applies the same graded logic: the higher the potential harm, the heavier the duties.
The Act matters to anyone who builds, sells, imports, distributes or uses AI in the EU market, including companies based outside the EU whose systems or outputs are used in the EU. Its obligations are being phased in between 2025 and 2028, and an amending regulation adopted in 2026, often called the Digital Omnibus on AI, moved the main high-risk deadlines. Teams need to know which category each system falls into, which role they play, and which dates apply.
What it is
The Act sorts AI into categories and assigns duties by role.
| Category | Examples | Main duties |
|---|---|---|
| Prohibited practices (Article 5) | Harmful manipulation or exploitation of vulnerabilities, social scoring, untargeted scraping of facial images to build databases, emotion recognition in workplaces and schools (with exceptions), certain biometric categorisation and real-time remote biometric identification for law enforcement (with narrow exceptions), and, from December 2026, systems generating non-consensual intimate imagery or child sexual abuse material | Not allowed |
| High-risk systems (Article 6, Annexes I and III) | Safety components of regulated products such as machinery or medical devices (Annex I); AI in biometrics, critical infrastructure, education, employment, essential services and credit, law enforcement, migration, justice and democratic processes (Annex III) | Risk management, data governance, technical documentation, logging, transparency to deployers, human oversight, accuracy, robustness and cybersecurity, quality management, conformity assessment, registration, post-market monitoring |
| Transparency obligations (Article 50) | Chatbots, AI-generated or manipulated content, deepfakes, emotion recognition and biometric categorisation | Inform people they are interacting with AI; mark synthetic content in machine-readable form; label deepfakes |
| General-purpose AI models (Chapter V) | Large models usable for many tasks | Technical documentation, information for downstream providers, copyright policy, training-content summary; extra duties for models with systemic risk |
| Minimal risk | Spam filters, most recommendation and analytics tools | No specific new obligations; voluntary codes encouraged |
Key roles are the provider, who develops a system or model and places it on the market under its name; the deployer, who uses a system under its authority; and importers and distributors. A company that substantially modifies a high-risk system or puts its own name on it can become a provider. All providers and deployers also have a duty to take measures for AI literacy among their staff, which the 2026 amendments softened.
Where it came from
The European Commission proposed the AI Act in April 2021, after a 2020 white paper on AI. The European Parliament and the Council reached political agreement in December 2023, adding rules for general-purpose AI models in response to the rapid rise of large language models. The regulation was signed in June 2024, published in the Official Journal on 12 July 2024, and entered into force on 1 August 2024.
Obligations apply in stages. Prohibitions and the AI literacy duty have applied since 2 February 2025. Obligations for general-purpose AI models, governance structures and penalties applied from 2 August 2025; models already on the market before that date have until 2 August 2027 to comply. The Commission published a General-Purpose AI Code of Practice in July 2025 as a voluntary way for model providers to show compliance. The general date of application, including the Article 50 transparency rules and the Commission's enforcement powers over general-purpose AI models, was 2 August 2026.
In November 2025 the Commission proposed a Digital Omnibus on AI to simplify the Act and adjust its timetable, partly because harmonised standards and guidance for high-risk systems were not ready. The amendments were adopted as Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026 and in force since 27 July 2026. Under it:
- Obligations for high-risk systems listed in Annex III apply from 2 December 2027.
- Obligations for high-risk systems that are safety components of products under Annex I apply from 2 August 2028.
- Providers of generative AI systems placed on the market before 2 August 2026 have until 2 December 2026 to meet the Article 50(2) marking duty.
- A new prohibition on AI systems generating non-consensual intimate imagery or child sexual abuse material applies from 2 December 2026.
- The AI literacy duty remains but is softened, with the Commission and member states tasked with supporting it.
- Simplifications for small and medium-sized enterprises are extended to small mid-cap companies.
- A new legal basis allows processing special category personal data where strictly necessary for detecting and correcting bias, under safeguards.
- The AI Office receives exclusive supervision over certain AI systems built on general-purpose models by the same provider and over AI in very large online platforms and search engines.
The obligations for general-purpose AI models were not postponed.
How it works
1. Inventory and classify
List every AI system and model you build or use. For each, check whether it is a prohibited practice, falls under Annex I or Annex III, triggers Article 50 transparency duties, or is minimal risk. Annex III systems can escape high-risk status under Article 6(3) if they do not pose a significant risk, for example performing a narrow procedural task, but this must be documented, and systems that profile people remain high-risk.
2. Determine your role
Decide whether you are provider, deployer, importer or distributor for each system. Duties differ sharply: providers carry most high-risk obligations, while deployers must use systems according to instructions, assign human oversight, monitor operation, keep logs and, in some cases, carry out a fundamental rights impact assessment.
3. Build the high-risk controls
Providers of high-risk systems set up a risk management system across the lifecycle, govern training, validation and test data for quality and bias, write technical documentation, enable automatic logging, provide instructions for use, design for effective human oversight, and test accuracy, robustness and cybersecurity. A quality management system ties these together.
4. Assess conformity and register
Before placing a high-risk system on the market, the provider performs a conformity assessment, internal for most Annex III systems and often with a notified body for product-related systems, issues an EU declaration of conformity, affixes the CE marking and registers the system in the EU database.
5. Meet transparency duties
Tell people when they interact with an AI system unless obvious, mark generated audio, image, video and text as AI-generated in a machine-readable way, and disclose deepfakes and certain AI-generated text published to inform the public.
6. Comply as a general-purpose model provider
Prepare technical documentation, give downstream providers the information they need, put in place a copyright policy, and publish a summary of training content using the Commission's template. Models with systemic risk, presumed above 10^25 floating-point operations of training compute, must also evaluate and mitigate systemic risks, report serious incidents and ensure cybersecurity.
7. Monitor and report
After deployment, providers run post-market monitoring and report serious incidents to authorities. National market surveillance authorities and the EU AI Office enforce the rules.
Key concepts and terms
| Term | Meaning |
|---|---|
| AI system | A machine-based system that infers from inputs how to generate outputs such as predictions, content, recommendations or decisions, with some autonomy. |
| Provider | The party that develops an AI system or model and places it on the market or into service under its name. |
| Deployer | The party using an AI system under its authority in a professional context. |
| High-risk AI system | A system listed in Annex III or a safety component of a product under Annex I legislation. |
| General-purpose AI model | A model with significant generality, able to perform a wide range of tasks. |
| Systemic risk | Risk from high-impact general-purpose models, presumed above a training-compute threshold. |
| Conformity assessment | The process showing a high-risk system meets the requirements. |
| Harmonised standard | A European standard whose use gives a presumption of conformity. |
| Fundamental rights impact assessment | An assessment some deployers must perform before using high-risk systems. |
| AI Office | The Commission body supervising general-purpose AI models and coordinating enforcement. |
| Post-market monitoring | Ongoing collection and analysis of performance data after release. |
| AI literacy | Staff knowledge and skills needed to use AI appropriately. |
A worked example
This scenario is illustrative. A software company based outside the EU sells a recruitment platform to employers in Germany and Spain. It has three AI features: a model that ranks applicants for a vacancy, a chatbot that answers candidates' questions, and a tool that drafts job adverts with a third-party language model.
The compliance team classifies each. Ranking applicants falls under employment in Annex III, so it is high-risk; the company is its provider. Because of the 2026 amendments, its high-risk obligations apply from 2 December 2027, giving the team about 13 months from November 2026. The chatbot triggers Article 50: candidates must be told they are talking to an AI, which already applies. The advert drafter uses a general-purpose model from another company; the model's provider handles model obligations, and the generated text needs marking duties assessed under Article 50.
For the ranking model, the team plans a risk management process, documents training data covering 1.2 million anonymised past applications, tests for disparities in selection rates across groups, writes instructions telling employers that a recruiter must review every shortlist, adds logging of each ranking with model version, and prepares technical documentation and a quality management system. They plan an internal conformity assessment, EU database registration and CE marking before the deadline, and appoint an authorised representative in the EU. Their employer customers, as deployers, will need to assign trained staff to oversee rankings and keep logs. Personal data in the training set also falls under the GDPR, so a data protection impact assessment runs in parallel.
Strengths and limitations
Strengths:
- A single framework across the EU single market.
- Risk-based: most AI faces no new rules, and duties scale with potential harm.
- Concrete requirements for documentation, data quality, oversight and robustness.
- Clear rules for general-purpose model providers, including transparency to downstream builders.
- Extraterritorial reach protects people in the EU regardless of where providers are based.
Limitations:
- Complex, with many definitions, exceptions and cross-references.
- Harmonised standards and guidance arrived late, which led to the 2026 postponements.
- Classification can be uncertain at the edges of Annex III.
- Compliance costs are significant for smaller providers, despite simplifications.
- Interaction with the GDPR, product safety and sector rules needs careful coordination.
- The timetable has changed once and remains subject to further guidance.
How it compares with alternatives
| Framework | Type | Relationship to the AI Act |
|---|---|---|
| GDPR | EU data protection law | Applies alongside whenever personal data is processed |
| NIST AI Risk Management Framework | Voluntary US framework | Useful method for risk management; not legally binding |
| ISO/IEC 42001 | Certifiable AI management system standard | Helps organise governance; supports but does not equal AI Act compliance |
| Council of Europe Framework Convention on AI | International treaty on AI and human rights | Broader principles; the AI Act is one way the EU implements them |
| AI governance programmes | Organisational policies and controls | The internal structure through which teams meet the Act |
Practical guidance and best practices
- Build an inventory of AI systems and models, with owners, purposes, users and markets.
- Classify each system and record the reasoning, especially for Article 6(3) exemptions.
- Decide your role for each system and check contracts with suppliers and customers for role allocation.
- Plan high-risk work against the amended dates: 2 December 2027 for Annex III, 2 August 2028 for Annex I.
- Meet the obligations that already apply: prohibitions, AI literacy, general-purpose model duties and Article 50 transparency.
- Integrate AI Act documentation with existing processes such as model cards, evals and model monitoring.
- Track harmonised standards, Commission guidelines and codes of practice as they are published.
- Coordinate with privacy teams so DPIAs and AI Act risk assessments share evidence.
- Review the classification whenever a system's purpose, users or model changes.
Common mistakes
- Assuming the postponement delayed everything; prohibitions, AI literacy, general-purpose model and transparency duties already apply.
- Treating a general-purpose model's compliance as covering the high-risk system built on it.
- Ignoring the Act because the company is based outside the EU.
- Classifying by technology rather than by intended purpose and context of use.
- Leaving human oversight as a checkbox without trained people, authority and time.
- Producing documentation once and never updating it as the model changes.
Frequently asked questions
When does the EU AI Act apply?
It entered into force on 1 August 2024 and applies in stages: prohibitions and AI literacy from 2 February 2025, general-purpose AI model duties from 2 August 2025, most other provisions including transparency from 2 August 2026, Annex III high-risk obligations from 2 December 2027 and Annex I high-risk obligations from 2 August 2028, following the 2026 amendments.
Does the AI Act apply outside the EU?
Yes. It applies to providers placing AI systems or general-purpose models on the EU market, and to providers and deployers outside the EU where the system's output is used in the EU.
What are the penalties?
Fines for prohibited practices can reach 35 million euros or 7% of worldwide annual turnover, whichever is higher; most other infringements up to 15 million euros or 3%; supplying incorrect information up to 7.5 million euros or 1%. Lower caps apply to SMEs, and separate fines apply to general-purpose model providers.
What changed with the Digital Omnibus on AI?
Regulation (EU) 2026/1744, in force since 27 July 2026, moved high-risk deadlines to 2 December 2027 and 2 August 2028, gave a grace period to 2 December 2026 for marking duties on existing generative systems, added a new prohibition, softened AI literacy, extended SME relief to small mid-caps and allowed special category data processing for bias correction under safeguards.
Is a chatbot high-risk?
Not by itself. A general chatbot usually has transparency duties under Article 50. It becomes high-risk only if used for a purpose listed in Annex III, such as deciding access to essential services.
Do open-source models have obligations?
Open-source general-purpose models released under free licences have reduced obligations, mainly a copyright policy and a training-content summary, unless they present systemic risk, in which case full obligations apply.
Glossary
- AI Act: Regulation (EU) 2024/1689 on artificial intelligence.
- Digital Omnibus on AI: Regulation (EU) 2026/1744 amending the AI Act.
- Annex III: the list of high-risk use areas.
- Annex I: EU product legislation whose safety components can be high-risk AI.
- GPAI: general-purpose AI.
- CE marking: the mark showing conformity with EU requirements.
- Notified body: an independent organisation designated to assess conformity.
- AI Office: the Commission's AI supervision body.
- Deepfake: AI-generated or manipulated content resembling real people or events.
- Code of practice: voluntary guidance used to demonstrate compliance.
Summary
The EU AI Act regulates AI by risk: it bans a set of practices, imposes extensive requirements on high-risk systems, adds transparency duties for chatbots and synthetic content, and sets obligations for general-purpose AI model providers. It entered into force on 1 August 2024, with prohibitions applying since February 2025, model duties since August 2025 and transparency rules since August 2026. The 2026 Digital Omnibus on AI, Regulation (EU) 2026/1744, moved high-risk obligations to 2 December 2027 for Annex III systems and 2 August 2028 for Annex I systems and made other targeted changes. Teams should inventory and classify systems, clarify roles, meet the duties already in force and use the extra time to build documentation, data governance and human oversight properly. This page is general information, not legal advice.
Related
It belongs to Security and governance in my stack, under MLOps and AI infrastructure. Others in the same group: