Everyone Thinks the EU AI Act Got Pushed to 2027. One Deadline Still Lands December 2, 2026, and Breaking It Costs €15 Million
The confusion is understandable. Two different clocks got set on two different parts of the same law, and only one of them moved. It's easy to see a headline about the AI Act being delayed and assume the whole thing slid a year. It didn't.
What actually happened: EU lawmakers reached a political deal in May 2026 to ease the law's rollout for 'high-risk' AI systems, things like hiring software, credit scoring, and medical devices, after companies and even some member states argued the original August 2026 start date left too little time to comply. That deal became Regulation (EU) 2026/1744, signed July 8, 2026, and in force since July 27, per the European Council's press materials on the agreement.
What actually moved, and what didn't
| AI Act provision | Covers | Deadline |
|---|---|---|
| Annex III high-risk obligations | Hiring tools, credit scoring, medical devices, and other 'high-risk' AI systems | Pushed to December 2, 2027, per Gibson Dunn. |
| Annex I product-safety component | AI embedded in regulated products (machinery, medical devices, toys) | Pushed to August 2, 2028, per Gibson Dunn. |
| Article 50(1), (3), (4): chatbot, emotion-recognition, and deepfake/text disclosure | Any provider or deployer of an interactive AI system, emotion-recognition tool, or deepfake/public-interest text generator | No extension. Already in force since August 2, 2026, per the EU's Article 50 guidance. |
| Article 50(2): machine-readable marking of synthetic content | Generative AI systems producing audio, image, or video output | August 2, 2026 for new systems. Systems already on the market before that date get until December 2, 2026. |
That last row is the one worth circling. If you launched a generative image, audio, or video tool before August 2, 2026, you have until December 2 to make its output machine-detectable as AI-made. If you launched after that date, the obligation already applies to you and has since day one.
The numbers
Dec 2, 2026
when the Article 50(2) grace period ends
for generative AI systems already on the market before August 2, 2026, per the Cloud Security Alliance's research note.
€15M
maximum fine for an Article 50 breach
or 3% of a company's total worldwide annual turnover, whichever is higher, per the EU's own guidance.
Dec 2, 2027
the deadline that actually moved
high-risk system obligations under Annex III, a full year later, per Gibson Dunn.
0
extensions granted to Article 50(1), (3), or (4)
chatbot disclosure, emotion-recognition notice, and deepfake labeling have applied since August 2, 2026 with no change, per the EU's transparency guidance.
Why 'machine-readable marking' isn't the same as a visible watermark
The law is deliberately technology-neutral. It requires output to be marked in a way a machine can detect as AI-generated, not that a human sees a logo stamped in the corner. Watermarking is the most common way providers are meeting that bar for audio, image, and video, but embedded metadata and cryptographic provenance signing (the same idea behind the C2PA content-credentials standard) satisfy it too, per the EU's Article 50 transparency guide. The Commission has said explicitly that no single method is considered sufficient on its own, so most providers are expected to combine at least two.
If you actually run one of these tools
- 1Check your launch date. If your generative image, audio, or video tool was live and available in the EU before August 2, 2026, the December 2 deadline is yours. If it launched after that date, you should already be compliant, not counting down to anything.
- 2Chatbot and deepfake disclosure already applies to you regardless. Article 50(1), (3), and (4) never had a grace period. If your product hasn't been telling EU users they're talking to an AI, or hasn't been labeling synthetic media as such, that's already a live violation, not a future one.
- 3Don't rely on one signal. A single visible watermark that a user can crop out, or metadata that a re-upload strips, likely won't satisfy 'machine-readable' on its own. Layer at least two methods.
- 4Watch February 2, 2027 too. Signatories to the EU's voluntary Transparency Code of Practice, finalized in June 2026, separately committed to stand up an interoperable watermark-detection mechanism by that date, a second deadline that can sneak up on teams who thought December 2 was the finish line.
Did the EU AI Act get delayed to 2027?
Only part of it. The Digital Omnibus pushed the 'high-risk' system rules (Annex III) to December 2, 2027, and a product-safety component (Annex I) to August 2, 2028. It left Article 50's transparency and marking rules alone.
What is the December 2, 2026 deadline, specifically?
It's the end of a four-month transitional window under Article 50(2) for generative AI systems that were already on the market before August 2, 2026. Those systems must have machine-readable marking on their synthetic output by December 2, 2026.
Does this apply to a chatbot that just answers in text?
The chatbot-disclosure duty under Article 50(1), telling a user they're interacting with AI, already applies and was never delayed. The December 2 grace period specifically covers marking synthetic audio, image, and video output under Article 50(2).
What's the penalty for missing it?
Up to €15 million or 3% of the company's total worldwide annual turnover, whichever figure is higher, per the EU's own transparency-rules guidance.
Does a visible watermark satisfy the requirement?
It can be part of the solution, but the law requires the mark to be machine-detectable, not just visible to a person, and the Commission has said no single technique is considered sufficient alone. Most providers combine watermarking with embedded metadata or a provenance-signing standard like C2PA.
Try the tools ✦
Free browser tools that never upload your files.