Your AI Act Marking Deadline Is 2 December 2026, Not 2027. The Delay Covered High-Risk, Not Article 50.
The Digital Omnibus pushed high-risk obligations to December 2027. Article 111(4) gave generative systems already on the market until 2 December 2026.
Most teams read one headline about the Digital Omnibus in July and filed the AI Act under 2027 business. The high-risk rules did move. Annex III standalone systems now fall due on 2 December 2027, and product-embedded Annex I systems on 2 August 2028. Transparency did not move. If your product was generating synthetic text, images, audio or video before 2 August 2026, your AI Act marking deadline is 2 December 2026, which is three weeks from today.
The obligation is short to state and awkward to satisfy. Article 50(2) requires outputs to be marked in a machine-readable format and detectable as artificially generated or manipulated, using technical solutions that meet four named criteria. No marking technique available in 2026 clears all four at once. Where each one breaks is well documented, and it matters less than it first appears, for a reason most compliance summaries skip.
Why the AI Act marking deadline for existing systems is 2 December 2026
The Digital Omnibus on AI, Regulation (EU) 2026/1744, entered into force on 27 July 2026 after Parliament adopted it on 16 June and the Council on 29 June. Its headline effect was to shift the high-risk application dates in Article 113. Annex III systems moved from 2 August 2026 to 2 December 2027, and Annex I systems to 2 August 2028. A clause-by-clause reading of the amendments shows how narrow the rest of it was.
Chapter IV, which carries Article 50, was not among the provisions moved. It applied from 2 August 2026 and has applied ever since. The prohibitions in Article 5 have run since February 2025, the AI literacy duty in Article 4 likewise, and the general-purpose model obligations kept their own schedule.
One transition did get written in. Article 111(4), as amended, says that providers of systems generating synthetic audio, image, video or text that were placed on the market or put into service before 2 August 2026 must take the necessary steps to comply with Article 50(2) by 2 December 2026. It is the only Article 50 duty with a transitional period attached, and it is grandfathering for an existing population rather than a delay for everyone.
The same date carries something else. The new prohibitions added to Article 5 on intimate-image and child sexual abuse material generation also begin to apply on 2 December 2026, under the third paragraph of Article 113. Those sit in the highest fine tier.
What Article 50(2) actually requires
Article 50(2) binds providers, including providers of general-purpose AI systems. It is distinct from Article 50(4), which binds deployers and is addressed to people: disclosing deepfakes and AI-generated text on matters of public interest to the audience that sees them. The provider duty is machine-to-machine. Something automated has to be able to read the output and tell.
The permitted techniques are open. Watermarks, metadata identifications, cryptographic provenance methods, logging and fingerprints all qualify, singly or in combination, provided the overall solution is machine-readable. The four criteria are that it be effective, interoperable, robust and reliable.
Three qualifiers do a lot of work. The duty applies as far as technically feasible. It takes account of the specificities and limitations of different content types. And it is measured against the generally acknowledged state of the art, which, as Stibbe reads it, means proven and commercially available technology rather than a research prototype. Implementation cost informs the proportionality analysis without becoming an exemption on its own.
The carve-outs are narrower than teams hope. A system performing an assistive function for standard editing is outside the duty, as is one that does not substantially alter the input data or its semantics, as are systems authorised by law for the detection, prevention, investigation or prosecution of criminal offences. Grammar correction is assistive. A model rewriting a paragraph is not.
The fine tier is set by Article 99. Breaches of Article 50 reach €15 million or 3% of total worldwide annual turnover, whichever is higher. For SMEs and start-ups, Article 99(6) inverts that to whichever is lower. National market surveillance authorities enforce it.
No single technique clears all four criteria
This is the part that vendor material tends to skate over. Each available approach is strong on some criteria and documented as weak on others.
| Technique | Survives handling | Cross-vendor detection | Where it breaks |
|---|---|---|---|
| C2PA signed metadata | Poor | Strong | Stripped on upload by major platforms; a screenshot removes it entirely |
| Imperceptible watermark | Moderate | Weak | Detection error rises after repeated marking; forgeable; no defence against a white-box attacker |
| Perceptual fingerprint plus logging | Moderate | Weak | Requires a query against the provider’s own index to resolve |
| Visible label alone | Poor | None | Not machine-readable, so it does not satisfy Article 50(2) by itself |
Signed metadata is cheap, interoperable and well specified, and it dies the moment content meets an indifferent intermediary. Platforms routinely strip metadata on upload, and a screenshot discards it completely. The coalition behind the standard says plainly that credentials can be stripped, that trust depends on who holds the signing certificate, and that the standard does not judge whether content is true.
Imperceptible watermarking survives re-encoding and cropping considerably better, which is why it is the usual answer to metadata stripping. The research record is also clear about its limits: detection error grows when content is marked repeatedly over its life, marks can be forged, and none of them holds against a determined adversary with white-box access.
Fingerprinting moves the authority to the provider’s own index. That is accurate and controllable, and it is the least interoperable of the three, because anyone wanting to check has to ask you. Pairing signed metadata with an imperceptible watermark is the configuration the Commission’s own code settles on, and the reason is that neither one is sufficient.
The gap the Act leaves open: nobody is forbidden from removing your mark
Here is the asymmetry that changes how the deadline should be approached. The AI Act does not expressly prohibit a third party from removing, altering or overwriting a machine-readable mark before content is disseminated. Copyright law does prohibit knowingly removing rights management information. The contrast is sharp enough that commentators have flagged it as a drafting gap rather than a deliberate silence.
The practical consequence is that the duty attaches to what a provider emits, and keeps emitting as techniques improve, rather than to whether the mark is still present when a regulator finds the content. The feasibility and state of the art qualifiers carry that weight. A provider who marks competently at generation time and whose mark is later stripped by a platform is in a different position from one who never marked at all.
The Commission can close the gap. Article 50(7), as amended by the Omnibus, has it encourage codes of practice on detecting, marking and labelling AI-generated content, and allows an implementing act setting common rules where a code proves inadequate.
What a defensible implementation looks like before 2 December
- Inventory every system that generates synthetic output, and establish for each one whether it was placed on the market or put into service before 2 August 2026.
- Emit two marks in the same pipeline: cryptographic provenance in signed metadata, and an imperceptible watermark in the content itself. One gives cross-vendor readability, the other survives handling.
- Publish a detection path. A mark that no third party can check is not detectable in any sense a regulator will accept.
- Cover text, not only images. Text is explicitly in scope and is where most inventories come up empty.
- Write down the proportionality reasoning while you are making the decisions, not afterwards.
- Re-read the exemptions against what your features actually do, rather than how they are described in marketing copy.
The Code of Practice is the benchmark you will be measured against
The Commission published the final Code of Practice on marking and labelling AI-generated content on 10 June 2026, ahead of the transparency duties taking effect. It is voluntary, and it operationalises exactly the two duties with the most technical content: provider marking under Article 50(2) and deployer labelling under Article 50(4).
It ships three ready-to-use EU icons, covering fully AI-generated content, AI-modified content, and a basic icon with an interactive second layer. It remains open for signature, with adequacy being assessed by the Commission and the AI Board.
Signing it immunises nobody. It can operate as a mitigating factor when an administrative fine is set, and more usefully, it is the clearest written account of what a regulator currently thinks an adequate solution looks like. For a team with three weeks, that is the document to design against.
After December
The harder question arrives once the deadline passes. Marking obligations assume a chain of custody that platforms, screenshots and re-encoding break by default, and the Act regulates only one end of that chain. Whether the next revision extends a duty to the intermediaries handling the content, as copyright law already does for rights management information, will decide whether any of this marking still carries meaning by 2028.
Frequently asked questions
Related reading
13 November Opens DPDP Consent Manager Registration. It Is Not Your Compliance Deadline.
Vendor blogs call 13 November 2026 a DPDP deadline. It is the date consent managers can register with the Board. Here is what changes, what does not, and what to do this month.
Stamp Duty on an Arbitral Award Is Due on Signing, Delhi High Court Says. Karnataka Disagrees.
Delhi High Court held that an arbitral award is chargeable to stamp duty when signed, and that a late deposit does not cure the defect. Karnataka reads the timing differently.
PAdES Signature Levels Explained: Why B-B Verification Breaks When the Certificate Expires
PAdES defines four signature levels, B-B, B-T, B-LT and B-LTA, each closing a way validation breaks over time. Most tools use only the weakest one. Here's what separates them.