Europe Dispatch Sunday, 22 November 2026
Europe Dispatch

Policy·European UnionExplained

The EU AI Act: what changes, when, and who it affects

The EU's AI Act sets rules for AI systems by risk level, bans nine practices and phases in obligations between 2025 and 2028, according to the European Union.

campaigners delivering a booklet to EU Parliament co-rapporteurs about the AI Act
Campaigners delivering a booklet to EU Parliament co-rapporteurs about the AI Act Ekō (formerly SumOfUs) / CC BY 2.0 / via Wikimedia Commons
Share on X WhatsApp

The AI Act is the first comprehensive legal framework on AI worldwide, aiming to foster trustworthy AI in Europe, according to the European Union. For particular uses of AI, the Act lays down rules based on risk that apply to those who develop and deploy such systems. The Act defines four levels of risk for AI systems.

The Act outlaws nine practices, among them harmful manipulation, exploitation of vulnerabilities, and assessing or predicting the risk of an individual committing a crime. Banned practices also include untargeted scraping of internet or CCTV material to build facial recognition databases and emotion recognition in workplaces and education. Prohibitions 1-8 took effect in February 2025, while prohibition 9, introduced as part of the AI Omnibus, takes effect in December 2026.

High-risk AI systems face strict obligations before market placement starting 2 December 2027, including risk assessment, data quality, logging, documentation, human oversight and robustness. Among the high-risk cases are safety components for critical infrastructure, scoring in education, surgery performed with robots, CV-sorting tools for recruitment and credit scoring.

Transparency rules require disclosure when humans interact with chatbots and labelling of deep fakes and certain public-interest text. The transparency rules of the AI Act come into effect in August 2026. The Act introduces no rules for minimal or no-risk AI, which covers most AI systems currently used in the EU, such as video games and spam filters.

Rules on General-Purpose AI models, covering transparency and copyright, became effective in August 2025. In July 2025 the Commission published three instruments on GPAI: guidelines on obligations, a voluntary Code of Practice, and a template for public summaries of training content. Technical documentation may be requested by the AI Office, which can also evaluate models, demand corrective measures and impose fines where compliance fails.

The AI Act took effect on 1 August 2024 and started applying on 2 August 2026, though certain exceptions exist. Prohibited practices and AI literacy obligations applied from 2 February 2025, and governance rules and GPAI obligations from 2 August 2025. Rules for high-risk AI in regulated products (Annex I) have an extended transition until 2 August 2028, and Annex III high-risk use cases until 2 December 2027.

The AI Omnibus simplification proposal was adopted on 19 November 2025, politically agreed on 7 May 2026 and entered into force on 27 July 2026. Under the Omnibus, rules for biometrics, critical infrastructure, education, employment, migration and border control apply from 2 December 2027, while rules for AI integrated into products such as lifts or toys apply from 2 August 2028.

The Omnibus introduced a prohibition on AI systems that create child sexual abuse material or sexually explicit and intimate content without consent, including nudification apps. It centralises oversight of AI systems built on GPAI models, strengthens the AI Office's powers and cuts governance fragmentation. Small mid-cap companies now benefit from simplified requirements, including lighter technical documentation, which were previously granted to SMEs, and access to regulatory sandboxes, among them one at EU level, is being widened so that more innovators can test their AI solutions under real-world conditions.

The Commission launched the AI Pact, a voluntary initiative to support future implementation and invite AI providers and deployers to comply with key obligations. A coordinated approach for Member States, businesses and public authorities is laid out in the July 2026 action plan on Cybersecurity and AI. The Commission will launch a call to boost EU evaluation capacity of AI models before market placement, expected to be operational by 2027, and the Commission and ENISA will create a blueprint for secure access to advanced AI systems and a secure testing platform for critical sectors.