EU AI Act deadlines: what applies when

Since 2 February 2025: prohibitions and AI literacy

The first wave: the prohibited practices of Article 5 (social scoring of people, manipulative techniques, emotion recognition at work and school, and the rest) and the AI-literacy duty of Article 4. Article 4 is the one small teams overlook — it applies to providers anddeployers with no size threshold, though it's a soft, proportionate duty: make sure the people operating AI on your behalf know what they're doing, and be able to show it. For a solo technical founder a short internal note is enough.

Since 2 August 2025: model providers

The general-purpose-model obligations of Chapter V and the governance machinery. This wave binds the Anthropics and OpenAIs of the world — not you as a downstream builder. If you're unsure which side of that line you're on, read provider or deployer.

Since 2 August 2026: the wave that hits products

The remainder of the Act applies — most importantly the Article 50 transparency obligations: chatbot disclosure, machine-readable marking of generated content, disclosure of published AI text. National market-surveillance enforcement starts, and anyone can lodge a complaint with an authority from this date. The Commission confirmed enforcement on 31 July 2026 — this date did not slip, despite recurring rumors.

2 December 2026: the marking grace period ends

One targeted transitional rule: AI systems already on the market before 2 August 2026 get four extra months — until 2 December 2026 — to implement the Article 50(2) machine-readable marking. Only the marking. The disclosure duties of 50(1) and 50(4) applied immediately, and content generated before August 2026 never needs retroactive labels.

The AI Omnibus moved high-risk to December 2027

The AI Omnibus — proposed November 2025, in force since 27 July 2026 — is the biggest schedule change. The Annex III high-risk obligations (the conformity assessments, registration, CE-marking apparatus) now apply from 2 December 2027 instead of August 2026, and product-embedded high-risk from August 2028. The Omnibus also added a prohibition on AI generating non-consensual intimate imagery, effective December 2026, and extended SME simplifications. Source: the Commission's AI framework page.

The dates people keep getting wrong

• "August 2026 is the high-risk deadline." It was, until the Omnibus. High-risk is now December 2027 — but August 2026 remains the transparency deadline, which is what actually touches most products.
• "There's a general grace period."There isn't. The only transitional relief for products is the 50(2) marking grandfathering to December 2026, and only for systems on the market before August 2026.
• "Old AI content has to be labeled retroactively." No — content generated before the obligations applied is out of scope.
• "Article 4 starts with the rest in 2026."AI literacy has applied since February 2025. It's cheap to satisfy; do it now.

Not sure which waves hit your product? The self-check sorts it out in eight questions.