
On 2 December 2026, providers of generative AI systems already on the EU market must begin marking what those systems produce so that software can detect it. Article 50 of the AI Act applied from 2 August, and a limited grace period covers earlier systems for this marking duty alone. A voluntary Code of Practice on transparency of AI-generated content, drafted by independent experts, has been assessed as adequate by the European Commission and the AI Board, which brings together Member State representatives. About 190 organisations had signed it by the end of July.
This is a pre-mortem: a check, before the outcome is known, on where a programme is most likely to fail. It asks five fixed questions, uses only public documents and judges the plan on the page, not the people behind it.
The Bet
The bet is that a legal duty, a voluntary Code and a grace period will put marking into products on time. Article 50(2) requires providers to mark synthetic audio, image, video and text in a machine-readable format so that it can be detected. The Commission’s guidelines set the scope, and the Code’s first section commits signatories to marking and detection that is effective, interoperable and reliable, as far as technically feasible. Signatories can rely on it to demonstrate compliance across Member States. Providers that do not sign must show their own measures are adequate, and the Commission says they may face more requests for information. The Code and the guidelines both appeared before the dates they serve.
The Assumption
The assumption is that marks can be made reliable enough to detect, and that detection will work across providers. The qualifier “as far as technically feasible” sits in the Code itself, and the Commission’s Q&A says the state of the art and the cost of implementation must be considered. Interoperability is still being built. The signatories’ marking and detection task force, set up on 30 September, lists among its objectives advancing the state of the art, standardising interfaces and establishing public benchmarks. Its page lists an interoperability working group meeting on 23 October, with minutes pending when checked on 14 November. Marking is a duty from 2 December. Detection across providers is still a work programme.
The Sequence
The sequence puts the duty ahead of the shared tools. The Code was published on 10 June and Article 50 applied on 2 August. Systems placed on the market after that date have had to comply from the start, and older systems have until 2 December. Over 90 signatories established the task forces on 30 September, nine weeks before the grace period ends. The enforcement powers of the AI Office, the Commission’s unit for AI, and the national authorities have applied since 2 August, so supervision of newer systems began two months before the task forces were established.
The Pager
Enforcement mainly sits with national market surveillance authorities, the bodies each Member State designates to check products on its market. The AI Office supervises systems built on a general-purpose model when the same entity provides both, and systems integrated into very large online platforms and search engines. Fines can reach 15 million euros or 3 per cent of worldwide turnover. The Commission notes that providers who do not sign will be assessed individually by different authorities. The Commission’s list of market surveillance contact points, updated on 9 November, marks 11 of 27 Member States as awaiting a final national designation decision, and shows no entry for three.
The Proof
Eighteen months from now, around May 2028, the test is whether people can tell which content is AI-generated without trying several detectors. That is the stated goal of the interoperability working group, and it is the right one. No figure or date is attached to it in the materials reviewed. The task force names public benchmarks and red-teaming exercises among its objectives, and the AI Office will review the Code at least every two years. What detection rate, on which kinds of content, the Commission will accept has not been stated. Nor do the materials reviewed give a baseline for how much AI-generated content is marked today.
Verdict
If the AI Office publishes a detection benchmark with a pass mark, and signatories report against it, then the marking duty can be judged on results and the Code can claim credit when it meets them. If the national authorities publish how they will check marks, then providers that chose not to sign will know the standard they are held to. A common way to read marks, delivered by the interoperability working group, would let anyone check content with one tool. The Commission has already supplied the rulebook, the guidelines and an adequacy finding. A marking duty is only as strong as the detector that reads the mark.


