The AI Act’s Marking Grace Period Ends on 2 December. If You Shipped After August, You Never Had One.

Ask a development team when the EU AI Act’s watermarking obligation kicks in and most will say 2 December 2026. It is in every compliance newsletter, it is three months away, and it feels like comfortable runway. For most of the products our clients are shipping right now, it is the wrong date entirely.

2 December is not the start of the marking obligation. It is the end of a four-month grace period granted to systems already on the market before 2 August 2026. If you shipped a generative feature after that date, you have been inside the obligation since the moment you placed it on the market. December is not your deadline. It is the date your last plausible excuse expires.

TL;DR

• Article 50 transparency obligations became applicable on 2 August 2026. Only systems placed on the market before that date got the four-month grace period ending 2 December 2026.
• Anything you shipped after 2 August 2026 owes machine-readable marking from the point of market placement, with no extension.
• The Digital Omnibus, Regulation (EU) 2026/1744, delayed the high-risk deadlines to 2027 and 2028. It left Article 50 alone.
• If your product carries your name or trademark, you are the provider under Article 3(3) and the marking duty is yours, not your model vendor’s.
• The hard part is not embedding the mark. It is stopping your own image pipeline, CDN and social card generator from stripping it, and hosting the free public detection the Code of Practice expects.
• Exposure under Article 99 runs to €15m or 3% of worldwide turnover, whichever is higher. For SMEs the calculation inverts to whichever is lower.

What the Omnibus actually moved

A lot of teams stopped reading the AI Act news cycle after the summer’s Digital Omnibus headlines and came away with a general impression that everything had been pushed back. That is half right in a way that is worse than being wrong.

Regulation (EU) 2026/1744 entered into force on 27 July 2026 and did move real deadlines: Annex III high-risk classification slipped to 2 December 2027, Annex I to 2 August 2028. If you are building recruitment scoring or credit decisioning, you genuinely bought time.

Article 50 was not touched. The transparency obligations became applicable on 2 August 2026 as originally scheduled. The only concession was a narrow transitional window for the Article 50(2) marking duty, and it was scoped to legacy deployments: generative systems already placed on the market before 2 August 2026 have until 2 December 2026 to retrofit machine-readable marking.

Read that scoping the way a market surveillance authority will. The grace period exists because retrofitting provenance marking into a running system is genuinely difficult, which is an admission that this is engineering work rather than a policy document. It is equally an admission that new builds have no such defence, because you could have designed it in.

You are probably the provider

The most common mistake we see is a team classifying itself as a deployer because it does not train models: we call an API, the model is somebody else’s, the marking is somebody else’s problem.

Article 3(3) does not care. A provider is a party that develops an AI system, or has one developed, and places it on the market or puts it into service under its own name or trademark. Wrap a hosted model in your product, put your logo on the output, and the Article 50(2) marking obligation attaches to you.

That changes who owns the failure. Your upstream vendor may well mark its outputs, but that mark is in their format, verified by their detector, and governed by their roadmap. You are discharging your own regulatory duty through a third party’s implementation detail, usually without a contract that says so, and frequently while your own pipeline destroys the mark in the next step.

The mark dies in your own pipeline

The Code of Practice on Transparency of AI-generated Content, published in final form by the Commission on 10 June 2026, endorses three mechanisms: cryptographically signed metadata of the kind C2PA Content Credentials produce, imperceptible watermarking embedded in pixels, audio waveforms or token distributions, and optional perceptual fingerprinting.

Signed metadata is the cheapest to adopt and the easiest to lose. It is a sidecar, and every stage of a modern asset pipeline is built to throw sidecars away for the sake of payload size:

• Image resizing libraries strip metadata by default unless told otherwise.
• CDN image optimisation re-encodes and discards everything non-essential.
• Upload paths that normalise formats produce a new file with a clean header.
• Open Graph and social card generators composite a fresh image from scratch.
• Third-party surfaces re-encode aggressively on the way in.

The obligation attaches to the output you place on the market. If your generation step writes a compliant Content Credential and your delivery step deletes it, you have built compliance and then removed it before anyone could see it. The Code is explicit that providers must not intentionally remove markings and should preserve existing provenance when generated content becomes an input to new generation. Nobody in that sentence intended anything. The default configuration did it.

This is why imperceptible watermarking is not an optional extra alongside metadata. It is the layer that survives the pipeline you did not audit. Auditing that pipeline is bounded, concrete work: generate an asset, walk it through every transformation between model and user, check for the mark at each hop. It takes an afternoon, most teams have never run it, and the result is usually bad.

The detection endpoint nobody budgeted for

Here is the part of Section 1 the summaries skip. Marking is half the duty. Outputs must also be detectable, and the Code expects providers to offer detection freely and publicly, via an open specification, downloadable software or a cloud API, with interoperability expectations landing in early 2027.

That is not a compliance artefact, it is a product surface. It needs a hostname, uptime, documentation, versioning, abuse handling and an owner, for as long as the content you generated is in circulation, which is forever. It is the sort of obligation waved through in a compliance workshop that lands on a backend team in January with no budget line and no rota.

Adopt an established open specification and its published detection interface and this collapses from a service you operate into a standard you conform to. Make that call deliberately now, rather than discovering it after shipping a bespoke watermark with no public detector.

Text remains genuinely unsolved

Marking text so the mark survives copy, paste and light editing is an open research problem, and pretending otherwise helps nobody. Article 50(2) asks for effectiveness and robustness “as far as technically feasible”, and that qualifier is doing real work for text. What it is not doing is excusing you from the rest. The Article 50(4) deployer duties are live with no grace period at all: disclose deepfakes, and disclose AI-generated or manipulated text published to inform the public on matters of public interest, unless it had genuine human review with a named person or organisation holding editorial responsibility. Superficial sign-off does not qualify. If you run a content operation with an AI drafting step, that is a workflow question with an audit trail attached, and it was due a month ago.

What to do before December

1. Classify each product in writing, with a name and a date. Provider or deployer, and if provider, placed on the market before or after 2 August 2026. That line decides whether you have twelve weeks or are already late.
2. Run the pipeline walk. One generated asset, every hop from model to user, check for the mark at each stage. Record where it dies.
3. Decide metadata plus watermarking, not metadata alone. Signed provenance for the clean path, an embedded signal for everything else.
4. Pick your detection story. Conform to a published specification with an existing detector, or accept you are operating a public service indefinitely.
5. Get marking into vendor contracts. What does your model supplier mark, in what format, with what detector, and what notice do you get if that changes?
6. Write the Article 50(4) editorial rule down. Who reviews AI-drafted public-facing text, what “review” means, and where the record lives.

The Code of Practice is voluntary and does not confer the formal presumption of conformity a harmonised standard eventually will. What it confers is the best available evidence of good faith, which is what a market surveillance authority will weigh. Ireland has taken a distributed approach to enforcement, designating competent authorities across existing sectoral regulators with the CCPC folding the AI Act into its administrative sanctions regime, so the conversation is likelier to arrive from a regulator you already deal with than from Brussels.

This is the fourth regulation in eighteen months to reach our clients as a specification rather than a policy document, and the pattern holds: the compliance industry sells the documentation, while the work sits in your build pipeline, your vendor contracts and your on-call rota. REPTILEHAUS builds and audits AI features for teams shipping into a regulated market, so if you want someone to run the pipeline walk, make the metadata and watermarking call, and get it done before December rather than during it, get in touch.

📷 Photo by David Nitschke on Unsplash