RecodeAI Explore AI Intelligence

AI Governance

The EU AI Act's watermarking grace period ends on 2 December

The headline this year was delay. The obligation arriving in ten days was never delayed, and it applies to systems that were already live in August.

2 min read

In brief

Widely reported delays to the EU AI Act covered high-risk obligations, not transparency. Article 50 has applied since 2 August 2026, and the transitional grace period for marking synthetic content on systems already on the market expires 2 December 2026 — ten days from publication.

What happened

This story is framed differently from the others in this edition, and deliberately so. We found no European Commission action in November 2026 on AI Act transparency. What we found instead is a deadline most organisations appear to believe was postponed, arriving in ten days.

The Digital Omnibus on AI — Regulation (EU) 2026/1744 — was published in the Official Journal on 24 July 2026 and entered into force on 27 July. It moved Annex III high-risk obligations to 2 December 2027 and Annex I obligations to 2 August 2028. It did not defer Article 50.

Article 50 transparency duties have applied since 2 August 2026: disclosing that a system is artificial intelligence when it interacts with people, marking synthetic content, and informing people about emotion recognition or biometric categorisation. For systems already on the market on 2 August, the Article 50(2) marking requirement carries a transitional period that expires 2 December 2026. Sources describe that period as three or four months; they agree on the date.

Why it matters

“The AI Act has been delayed” was the sentence most organisations took away from the summer. It was true of the part that was hardest to comply with and false of the part arriving soonest. That gap between the headline and the regulation is itself the risk.

Business and industry implications

The exposure is concentrated in systems deployed before August that nobody has revisited since — a chatbot on a support page, generated imagery in a campaign, synthetic voice in an IVR. These were in scope from August and lose their transitional cover in December. They are also, typically, owned by marketing or service functions rather than by whoever tracks AI regulation.

Compliance here is unusually tractable: disclosure and marking, not a conformity assessment. The difficulty is inventory. Most organisations cannot currently list every AI-touching surface a customer meets.

Leadership perspective

Governance fails at the boundary between who deploys AI and who tracks the rules. Embed the question into ordinary technology and procurement decisions — does this touch a person, does it generate content, is it marked — rather than running it as a separate compliance exercise that only reaches systems someone remembered to flag.

These are law-firm and analyst readings, not Commission texts. For compliance decisions, work from the Official Journal text of Regulation (EU) 2026/1744 and Commission guidance directly.

Three questions for business leaders

  1. Can we produce, this week, a list of every customer-facing surface that uses AI or generates synthetic content?
  2. Which of those were live before 2 August 2026 and have not been reviewed since?
  3. Who owns this — and do they know the December date applies to them?