AIACT50 newsletter · week of 27 October 2026. The hardest part of Article 50 is still outstanding, and the clock on it is now measured in weeks.
In August, a lot of organisations closed the file on Article 50 of the EU AI Act. The transparency obligations became applicable and enforceable by national authorities on 2 August 2026, the chatbot-disclosure and deepfake-labelling duties went live, and compliance teams moved to the next item on the list. That was a reasonable thing to do. It was also incomplete, and the gap is about to matter.
The part of Article 50 still outstanding
The part of Article 50 that is hardest to build — machine-readable marking of generated output, so that a third party can detect that an image, a clip, or a block of text came from an AI system — was deferred for providers whose generative systems were already on the market before August. That deferral expires on 2 December 2026. As this goes out, that is roughly five weeks away. For a large share of enterprises running generative features today, the most technical requirement in the whole transparency regime is the one still outstanding, and the clock on it is now measured in weeks.
The Omnibus eased high-risk, not this
The confusion is understandable, because the surrounding news pushed in the opposite direction. The Digital Omnibus — published in the Official Journal in late July — was widely reported as the EU easing the AI Act, and for high-risk systems that is accurate: those obligations slipped to 2 December 2027 and 2 August 2028, and a range of procedural requirements were simplified. If your only exposure is high-risk, you did get more time.
For anyone who generates or publishes synthetic media, the same instrument reads differently. It left Article 50 transparency on its original timeline — no delay — so the 2 December marking deadline for existing generative systems did not move. And it added a new prohibition, under Article 5, on AI systems that generate certain categories of illegal content, effective the same day, with no transition window. We mention that prohibition only to make the calendar honest: it is a safety and moderation obligation, not a transparency one, and not something a marking tool discharges. But its effect on the calendar is real. For the synthetic-media segment specifically, 2 December 2026 now carries two distinct obligations at once, which makes it, for that group, the most demanding single date the Act has produced so far.
Marking converges. Verification fragments.
There is a quieter development underneath all this. The question of how to mark AI content is being settled by the model providers themselves: OpenAI and Google have both committed to Google's SynthID watermark alongside C2PA Content Credentials, with others following. That is genuine progress toward a common marking vocabulary. But the question of how anyone verifies a mark is moving the other way. Google's checker now wants a Google account; OpenAI's planned page will read only OpenAI's own content.
Verification is fragmenting into a set of gated, per-provider portals — the opposite of the free, public, interoperable check the EU's own Code of Practice describes.
Norway's news sector, meanwhile, is quietly trying to operationalise Content Credentials at ecosystem scale, with a target of 80% adoption — a reminder that the marking layer is becoming real infrastructure, which only sharpens the verification gap.
One thing worth doing this week
If there is one thing worth doing this week, it is unglamorous: take a single live output from your generation pipeline and check whether a machine could tell it was AI-generated. If it can't, 2 December is a build deadline, not a review date. The architecture the regulators are converging on — marking that survives, and a verification anyone can run — is buildable today. The scarce resource now is the five weeks.
Sources: Goodwin, Aug 2026 · Baker Botts, Sep 2026 · Paul Weiss, 27 Jul 2026 · ActuIA (Reg (EU) 2026/1744) · Cookiebot — 2 December deadline · AI Insiders — Omnibus & Article 5 · Gibson Dunn — high-risk delays · McCann FitzGerald — one month to go · heise — OpenAI & Google on SynthID + C2PA · Content Authenticity Initiative — Norway Project Reynir, 21 Oct 2026
Marking that survives, and a verification anyone can run:
Explore AI Act 50 →