AI Labelling Becomes a Product Feature: Watermarks, C2PA and the December 2 Deadline

On 1 September 2026, Anthropic released two new models, Claude Fable 5.1 and Claude Mythos 5.1. The part of the announcement that matters to anyone responsible for compliance is not in the benchmark table: text from these models carries a watermark from the provider, and files produced by code execution receive C2PA Content Credentials — provenance data that travels inside the file itself.

More interesting than the technical detail is the shift behind it. For two years, labelling was a filing-cabinet task: a duty you discharge by writing something alongside. It is now moving into the product — the provider ships the marking whether you ordered it or not. That makes what Article 50(2) of the AI Act has required since 2 August 2026 technically achievable for the first time. Which is precisely why it is worth looking closely at what this marking covers — and what remains your own obligation.

At a glance

• What applies: Article 50 of the AI Act has been applicable since 2 August 2026. The machine-readable marking under paragraph 2 binds providers; the disclosure duties under paragraphs 1 and 4 have no grace period.
• What follows: Since 1 September 2026, text output from Fable 5.1 and Mythos 5.1 carries a watermark, and files from code execution carry C2PA Content Credentials. Labelling is now a product feature.
• The catch: For systems placed on the market before 2 August 2026, the deadline for paragraph 2 expires on 2 December 2026 (the new Article 111(4)). The Code of Practice and the EU icon set are orientation, not a presumption of conformity.

What went into the product on 1 September

The facts, briefly: Claude Fable 5.1 (claude-fable-5-1) and Claude Mythos 5.1 (claude-mythos-5-1) appeared on 1 September 2026, both with a one-million-token context, at 10 US dollars per million input and 50 US dollars per million output tokens. According to the provider, Mythos 5.1 is the same model with reduced safety guardrails, reachable only through a vetted access programme for specialists.

Two lines matter for labelling. First: text from these models carries Anthropic’s text watermark; the provider had already explained how it works on 14 August 2026. Second: files produced by code execution receive C2PA Content Credentials — the provenance sits in the file, not in a notice you lose when forwarding it.

The law, read closely

Article 50 of the AI Act has been applicable since 2 August 2026. The Digital Omnibus changed nothing about it materially — only paragraph 7 was touched, the procedure for codes of practice. What was postponed is the high-risk regime, not transparency; I have set out the revised timetable separately. The four sets of duties stand under different conditions:

Set of duties (Art. 50 AI Act)	What is required	Grace period to 02/12/2026
Para. 1 — direct interaction	people must be able to tell they are talking to an AI system — a notice at the input or output field, for voice systems at the start of the conversation	no
Para. 2 — synthetic content	machine-readable marking by the provider: watermarks, metadata, content credentials	yes, for legacy systems only
Para. 3 — emotion recognition, biometric categorisation	disclosure to the persons concerned	no
Para. 4 — deepfakes, AI text on matters of public interest	labelling or disclosure — except where there is human review with named editorial responsibility	no

The most important sentence for planning is not in Article 50 but in the new Article 111(4): “Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, that have been placed on the market before 2 August 2026 shall take the necessary steps in order to comply with Article 50(2) by 2 December 2026.” So anyone who had a generative system on the market before that date has until 2 December 2026 for the marking — for everything else in Article 50, the deadline does not apply.

Directly relevant for website operators is the exception in Article 50(4), second subparagraph: the duty to disclose AI text on matters of public interest falls away where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for it. A transparency page plus documented review, in other words, rather than labelling every paragraph.

The Commission’s final guidelines on Article 50 have been available since 20 July 2026. They are not legally binding, but they are the yardstick supervisors will apply: the information must be clear and comprehensible and given at the latest on first interaction.

Where it falls short

Three objections belong here, and the first matters most in practice.

First: the provider’s duty is not your duty. The watermark addresses Article 50(2) — a provider obligation. Yours are different: making clear that your chatbot is a machine; labelling deepfakes; disclosing AI text on matters of public interest. Those duties have applied without any grace period since 2 August 2026. A provider that marks its outputs makes it easier, at most, for you to prove machine origin — it cannot take anything off your plate.

Second: the standard is qualified by feasibility, and the coverage is narrow. The Code of Practice asks providers for solutions that are “effective, interoperable, robust, and reliable” — “as far as technically feasible”. That qualifier is the honest part of the formula. What was actually covered on 1 September: a text watermark for two models from one provider, and content credentials for files from one execution path. What another model outputs, what someone copies out or paraphrases — the marking says nothing about any of it. And a marking is worth only as much as the ability to read it: anyone who wants to use it in a release process needs a tool for that.

Third: the Code and the icons are orientation, not proof. The Code of Practice on Transparency of AI-generated Content has been available since 10 June 2026, is regarded by the Commission and the AI Board as an adequate means, and had around 190 signatories by the end of July 2026; its EU icon set helps with the visual implementation. It does not create a presumption of conformity — the recitals of the amending regulation say so expressly: codes have “limited legal effect, and in particular do not grant a presumption of conformity”. Signing up is positioning, not a liability shield.

The marking travels only along the path it was built for. What runs beside it carries nothing — and that is where your own disclosure begins.

What this means for your business

First, work out which of the four duties actually applies to you. In the typical mid-sized case it is not paragraph 2 but paragraphs 1 and 4: the chatbot on your website and the texts you publish. For the chatbot case I have worked through transparency, liability and data protection separately — the effort is small, and the deadline for it passed on 2 August.

Go with a transparency page plus documented editorial review. That is the workable route through paragraph 4: a page that discloses how the content is produced, and a named person who holds responsibility. This website does it that way — see AI transparency. What is decisive is that the review genuinely happens: the exception attaches to the review, not to the claim that there is one.

Treat 2 December 2026 as a contract matter, not a DIY project. If you use generative AI in a product that was already on the market before 2 August 2026, your grace period ends on that day. You usually do not have to build the marking yourself — you have to know whether your provider supplies it and whether it is warranted. That is a question for the contract; which clauses belong in it is set out in AI contracts.

Conclusion

Two things are remarkable here, and they point in different directions. The good one: a duty that a year ago still sounded like unachievable bureaucracy has arrived in the product — without a change in the law, without an implementing act, simply because a provider built it in. The uncomfortable one: that very convenience invites the wrong conclusion. Your tool marking its output does not mean you have labelled anything.

The sober reading: the provider’s marking eases the technical side, it does not replace your own disclosure. And 2 December 2026 sits inconspicuously between the big dates — which is why it gets overlooked.

If you want to establish which of the four duties under Article 50 applies to your system, and how to evidence compliance, let’s talk. I read questions like these as a business lawyer and build the systems in question myself.

FAQ

Do I have to label AI-generated text on my website?

Only to a limited extent — and not paragraph by paragraph. Article 50(4) of the AI Act requires disclosure for AI-generated text on matters of public interest. The obligation falls away where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for it. For a company or law-firm website, the workable route is therefore a transparency page plus editorial review that genuinely takes place.

What happens on 2 December 2026?

That is the day the grace period for legacy systems ends. The new Article 111(4) of the AI Act gives providers of generative systems placed on the market before 2 August 2026 until then to implement the machine-readable marking required by Article 50(2). The deadline applies only to paragraph 2 and only to legacy systems — the disclosure duties under paragraphs 1 and 4 have applied without any grace period since 2 August 2026. Two new prohibitions in Article 5 also take effect the same day.

Does the provider’s watermark replace my own labelling?

No. The machine-readable marking under Article 50(2) is a provider obligation; a model provider who builds it in discharges its own duty, not yours. Your duties are different ones: making clear that a chatbot is a machine, labelling deepfakes, and disclosing AI text on matters of public interest. The provider’s marking can make it easier to show that a piece of content was machine-generated — it cannot take anything off your plate.

Is signing the Code of Practice enough to prove conformity?

No. The Code of Practice on Transparency of AI-generated Content is voluntary; the Commission and the AI Board assessed it as an adequate means, and by the end of July 2026 it had around 190 signatories. It expressly does not create a presumption of conformity — the recitals of the amending regulation speak of “limited legal effect, and in particular do not grant a presumption of conformity”. The EU icon set it contains helps with the visual implementation, but it is not a liability shield.

---

Sources — as of 02/09/2026

• Commission: guidelines on Article 50 (transparency of AI-generated content) — https://digital-strategy.ec.europa.eu/en/policies/guidelines-transparency-ai-generated-content
• Commission: FAQ on Article 50 — https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act
• Commission: Code of Practice on Transparency of AI-generated Content — https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content
• Article 50 of the AI Act, article text (German) — https://ai-act-law.eu/de/artikel/50/
• Regulation (EU) 2026/1744 (Digital Omnibus on AI), full text — including the new Article 111(4) — https://publications.europa.eu/resource/celex/32026R1744
• Anthropic (primary source): Claude Fable 5.1 and Claude Mythos 5.1 (2026-09-01) — https://www.anthropic.com/claude-fable-and-mythos-5-1
• Anthropic: How Claude’s text watermark works (2026-08-14) — https://www.anthropic.com/news/claude-text-watermark

Note: the date 20 July 2026 for the final guidelines comes from law-firm briefings; the Commission page shows only a last-update date. The publication date and signatory count of the Code of Practice are well sourced but not officially confirmed.

This article is general information and not legal advice in an individual case. As of 2 September 2026; the legal position on the AI Act continues to move, so please check the current state before making decisions.