EU AI Act watermarking requirements: what you must ship by December 2, 2026

The EU AI Act watermarking requirements have been in force since August 2, 2026. December 2, 2026 is not a start date - it is the end of a four-month grace period that covers one duty only: machine-readable marking of synthetic output by generative systems that were already on the EU market before August.

If you ship software with an AI feature to users in the EU, that distinction decides what you owe this quarter. The chatbot disclosure, the deepfake label and the notice on emotion recognition carried no grace period at all. Below: who owes what, what counts as marking, what is exempt, and the fines.

The four transparency duties, and who owes each one

Article 50 of the AI Act splits into four duties that attach to a function, not to a risk tier. The Commission's own FAQ divides them between providers - whoever develops an AI system and places it on the EU market under their own name - and deployers, who use a system under their own authority in the course of a professional activity. Purely personal, non-professional use is out of scope.

Two points in the Commission's transparency obligations FAQ are worth pinning down early. First, a provider outside the EU is covered when the system's output is used in the EU - there is no establishment test to hide behind. Second, you can be both at once: you are a provider of your own assistant and a deployer of the model underneath it.

What the December 2, 2026 date actually covers

The grace period exists because of the Digital Omnibus on AI, Regulation (EU) 2026/1744, which entered into force on July 27, 2026. That regulation pushed the main obligations for stand-alone high-risk systems in Annex III to December 2, 2027, and for AI embedded in regulated products under Annex I to August 2, 2028. It did not move Article 50.

What it did move is narrow. Systems placed on the EU market before August 2, 2026 get until December 2, 2026 to meet the marking and detection duty under Article 50(2) - and nothing else. The Commission's FAQ is explicit that the grace period applies only to that sub-obligation.

One more carve-out is worth knowing before you plan a backfill: content generated before August 2, 2026 does not have to be labelled retroactively, though the Commission encourages it where possible. So the December work is about the pipeline from here on, not your archive.

In our view the most common mistake right now is reading old commentary. Plenty of still-ranking pages were written while the Omnibus was a proposal and describe the whole AI Act timetable as slipping. The delay is real, it is law, and it never touched the disclosure duties.

EU AI Act watermarking requirements in practice: expect two techniques, not one

Article 50(2) is technology-neutral. It asks that outputs be marked in a machine-readable format and detectable as artificially generated or manipulated, where technically feasible - it does not name a watermark.

The practical standard comes from the Code of Practice on Transparency of AI-generated Content. Reading the final package, Paul, Weiss notes in an August 4, 2026 client memo that the Commission takes the view that no single technique currently meets the standard, so the Code expects signatories to combine at least two - for example digitally signed, tamper-evident metadata plus an imperceptible watermark - and not to let users strip the metadata. Per the same memo, a single technique is accepted for free-form text, which cannot carry metadata, and for closed physical products whose outputs cannot be shared online. Providers are also expected to offer a detection mechanism, with a further deferral to February 2, 2027 for making watermark detectors interoperable.

For the metadata half, the open standard with the widest industry backing is C2PA's Content Credentials, which its own site describes as an open technical standard for establishing the origin and edits of digital content, now at specification version 2.3, with Adobe, Google, Meta, Microsoft, OpenAI and TikTok among the eleven steering committee members. Signing your outputs with Content Credentials is not the same as complying with Article 50, but in our view it is the part of the job you can do with existing libraries rather than original research.

On the detection half, we'd buy rather than build. Hive AI, one of 194 tools in APIs & Integrations on Upfynd as of October 2026, sells AI-generated and deepfake detection models that return confidence scores for images and video, for audio in 10-second chunks, and for music, according to its content detection API page; pricing is not published and the page routes to a sales contact. Treat detection vendors as evidence-gathering for moderation and disputes, not as a substitute for marking your own output.

Your chatbot already has to say it is an AI

Article 50(1) asks providers to design interactive systems so that a person is informed they are dealing with an AI, unless that is obvious. The Commission's FAQ says the "obvious" exception is interpreted restrictively, which is the part that decides the edge cases: a product named "Assistant" does not clear it, and a sentence in your terms of service is not disclosure at all.

Per the Paul, Weiss memo, the finalized guidelines go further for agents - an AI agent must disclose both that it is artificial and the person on whose behalf it is acting - and raise the bar where children, older people or people with disabilities may be in the audience. If you build on this pattern, the 208 tools in AI Agents are worth scanning for how they surface that at first contact, because Article 50(5) requires the information to land clearly and distinguishably at the latest at the first interaction or exposure.

We'd treat this as cheap to fix and expensive to forget: a persistent label in the chat header is an afternoon of work, and it is the first thing a market surveillance authority can check from the outside.

Deepfakes and published AI text are the deployer's problem

If you use AI to create or manipulate image, audio or video content that looks plausibly real, you disclose it. Two details from the Commission FAQ decide most cases. A provider's machine-readable mark does not discharge your labelling duty - the marking and the label are separate obligations with separate owners. And where the work is evidently artistic, creative, satirical or fictional, disclosure must be limited so it does not hamper enjoyment of the work, which is a narrower relief than an exemption.

Published AI-generated text informing the public on matters of public interest also needs a label, unless it went through human review or editorial control with someone responsible for it. The FAQ says spell-checking alone does not qualify. If you run a programmatic blog or a localized help center for an EU audience, that is the sentence to read twice.

What is exempt from the marking duty

The Commission FAQ lists the carve-outs from Article 50(2). Marking is not required for:

• short sequences of numbers, symbols or letters
• source code
• outputs used exclusively for machine-to-machine processing with no human exposure
• closed-loop industrial and product-development outputs, unless they are the final output
• assistive functions for standard editing, or systems that do not substantially alter the input data
• certain business-to-business and industrial contexts, under conditions set in the guidelines

The Code of Practice adds a threshold for text that signatories can rely on: per the Paul, Weiss memo, text under roughly 200 tokens - about 150 words - is exempt from watermarking. On that wording, autocomplete, grammar fixes and format conversion sit outside the duty; a generated blog post does not.

The Code of Practice is voluntary; Article 50 is not

The Commission's Code of Practice page states the position plainly: adherence is voluntary, but the Article 50 transparency requirements are legal obligations either way. Section 1 covers providers on marking and detection; Section 2 covers deployers on labelling deepfakes and certain published content. The Commission concluded that the Code adequately covers the obligations in Article 50(2), (4) and (5), and says non-adherents must demonstrate compliance by other adequate means.

That is the trade. Signing gives you the only framework the Commission has endorsed for showing you comply; not signing means writing your own justification and defending it. Around 190 organizations had signed as of July 31, 2026, per the Commission announcement cited in the Paul, Weiss memo, and the AI Office's window for signing ahead of the August date closed on July 27, 2026.

We'd sign if you are a provider of a generative feature and skip it if you are only a deployer labelling occasional marketing assets - the deployer section asks for little you would not do anyway.

What the fines are, and why the cap is different for startups

Breaching Article 50 falls under the Act's general penalty tier for operator obligations: up to €15 million or 3% of total worldwide annual turnover, per the Commission's FAQ, which also notes that proportionality can apply to SMEs and small mid-caps. The direction of that cap is the part worth knowing. Under Article 99(6), for SMEs including start-ups each fine is capped at the lower of the two figures rather than the higher - so turnover is a ceiling for a small company, not a multiplier.

Enforcement sits mainly with national market surveillance authorities rather than the AI Office, which means your exposure follows the member states where your users are.

What we'd ship before December 2

In priority order, assuming a SaaS product with a generative feature and EU users:

1. Audit which of the four duties you owe, as provider and as deployer, and write down the answer with a date on it.
2. Put AI disclosure at first interaction in every conversational surface, including agents acting on a user's behalf.
3. Add signed, tamper-evident provenance metadata to generated media, and an imperceptible watermark alongside it where the format allows.
4. Expose a detection or verification endpoint so a recipient can check your marks.
5. Decide on the Code of Practice, and if you skip it, document the alternative means you are relying on.
6. Fix the deployer side: label published AI text aimed at an EU public audience, and record who holds editorial responsibility where you claim the review exemption.

The tooling for steps 1 and 5 is governance and documentation work, which is where Legal & Compliance - 80 products on Upfynd as of October 2026 - and the 383 tools in Operations overlap; our earlier SaaS launch checklist covers the chatbot and review side of a launch in the same spirit. Nothing here is legal advice, and the guidelines are long: read the Commission FAQ before you commit an implementation.

If you make a compliance or provenance tool that does part of this job, list it on Upfynd - the December 2 date is six weeks away, and Legal & Compliance is one of our thinner categories.