YARD Law Co. · Reviewed August 2026 · YARD Law Legal Team
On 14 August 2026, Anthropic explained a change that will gradually affect text generated by Claude: future Claude models will place an invisible watermark into their written output.
It is not a hidden Unicode character. It is not metadata attached to the document. And it does not insert a visible statement saying that a passage was written by Claude.
Instead, the watermark is produced during generation itself.
Anthropic uses a version of Google DeepMind's SynthID-Text technique. Large language models repeatedly choose the next token from several plausible alternatives. The watermark modifies the mechanism used to make some of those choices so that, across a sufficiently long passage, the resulting sequence contains a statistical pattern detectable with the appropriate key.
In that sense, the watermark is not attached to the finished text. The generated text carries the signal itself.
The immediate reason for the change is regulatory.
The transparency provisions of Regulation (EU) 2024/1689, the EU Artificial Intelligence Act, became applicable on 2 August 2026.
Article 50(2) requires providers of AI systems generating synthetic text, images, audio or video to ensure that their outputs are marked in a machine-readable format and can be detected as artificially generated or manipulated.
The Act does not prescribe one specific technology.
An AI provider may use watermarking, provenance technology or another sufficiently effective solution, provided that the mechanism satisfies the requirements of effectiveness, interoperability, robustness and reliability as far as technically feasible.
Anthropic has chosen statistical text watermarking.
Traditional digital provenance often relies on metadata.
Metadata can say that an image was created with a particular application or contain a cryptographically signed Content Credential. The weakness is that metadata can disappear when a file is converted, copied, screenshotted or processed by a platform.
A statistical text watermark takes a different approach.
The signal emerges from the words selected by the model themselves.
Copying and pasting the text therefore does not automatically remove the watermark.
Editing presents a more difficult question. Anthropic says that light editing may leave sufficient signal for detection, while a complete rewrite can remove it.
That immediately creates a new technical and legal problem: where does AI-generated content end and human-authored content begin?
This is the most important limitation.
Anthropic expressly states that its detector will establish only the likelihood that Claude was involved in producing or processing the text.
It cannot necessarily distinguish between:
Translations are particularly instructive. Anthropic says a translation produced by Claude will carry the watermark because Claude selects every word of the translated output.
The underlying ideas and original authorship may nevertheless remain entirely human.
A positive watermark result should therefore not automatically be treated as proof of plagiarism, AI authorship or absence of human intellectual contribution.
This distinction may become increasingly important for universities, employers, publishers and courts as provider-specific detection tools become available.
The technical watermarking obligation is only one part of Article 50.
The Act draws a distinction between the provider of the AI system and the deployer who uses it.
Article 50(2) is primarily a technical obligation imposed on providers such as Anthropic, OpenAI or Google.
The provider must build machine-readable marking and detection into the relevant AI system.
This is the obligation Anthropic's new watermark principally addresses.
Article 50(4) creates a separate obligation for professional users of generative AI.
Where AI-generated or AI-manipulated text is published for the purpose of informing the public about matters of public interest, the deployer must disclose that the text has been artificially generated or manipulated.
There is, however, an important exception.
Disclosure is not required where the content has undergone human review or editorial control and a natural or legal person assumes editorial responsibility for the publication.
A news organisation, law firm or professional publisher that genuinely researches, verifies, edits and assumes responsibility for the finished publication may therefore be in a different legal position from a website automatically publishing unreviewed AI output.
This produces an unusual consequence.
A professionally edited article may contain an Anthropic watermark because Claude participated in its preparation.
At the same time, Article 50(4) may not require the publisher to display a visible “AI-generated” label because the article has undergone genuine human editorial review and a person or organisation assumes responsibility for it.
There is no contradiction.
The watermark answers one question:
Was this AI system probably involved in producing the text?
The disclosure rule addresses another:
Must the publisher tell the reader that the publication was artificially generated or manipulated?
Those questions should not be conflated.
Article 50(2) itself contains an exception where an AI system merely performs standard editing assistance or does not substantially alter the user's input or its semantics.
This is also reflected technically in Anthropic's description of its system.
If Claude only corrects a few punctuation or grammar errors, there may be too few model-generated choices for a meaningful watermark to be detectable.
A substantial rewrite or translation is different.
The EU AI Act has broad territorial reach.
It applies to providers placing AI systems on the European market regardless of whether the provider itself is established in the EU. It can also apply to providers and deployers outside the Union where AI outputs are used within the EU.
Anthropic is therefore subject to the Regulation even though it is a US company.
More unusually, Anthropic has announced that its initial watermarking rollout will be global.
The company says it does not currently have a durable technical way to apply the watermark only to European users.
As a result, a Claude user in the United States, Asia or another non-EU jurisdiction can receive the same watermarked output.
The watermark does not contain the user's identity, organisation or conversation information.
The EU has also adopted a Code of Practice on Transparency of AI-Generated Content to help providers and professional users demonstrate compliance with Article 50.
Anthropic signed the provider section of the Code.
Other signatories include Google, OpenAI, Meta, Microsoft, Mistral, Cohere, Aleph Alpha and Synthesia.
That does not mean every provider will use the same watermark.
Google already uses its SynthID technology across several kinds of generated content, including text. OpenAI has publicly deployed provenance technologies including C2PA Content Credentials and SynthID for supported generated media.
Other providers may implement different technical combinations.
The regulatory obligation is the same; the technical architecture is not.
The AI Act's transparency obligations generally became applicable on 2 August 2026.
However, Regulation (EU) 2026/1744, the Digital Omnibus on AI, introduced a limited transition for generative AI systems that had already been placed on the EU market before that date.
Those systems must comply with Article 50(2) by 2 December 2026.
The wider effect of that Regulation on the AI Act timetable, including what applied on 2 August 2026 and what was deferred, is covered in our Bulgarian article AI Act: what took effect on 2 August 2026 and what was postponed.
Anthropic has therefore said that watermarking will also be rolled out to older Claude models during the coming months.
Failure to comply with Article 50 can trigger significant sanctions.
For undertakings, the AI Act provides for administrative fines of up to EUR 15 million or 3% of total worldwide annual turnover for the preceding financial year, whichever is higher.
For SMEs, special proportionality rules apply.
Financial penalties are not the only enforcement tool.
Market-surveillance authorities can require corrective action and, in serious or continuing cases, may restrict or prohibit an AI system from being made available, or require its withdrawal or recall from the market.
For a global AI provider, access to the European market is therefore part of the compliance risk.
Possibly, but its evidential meaning needs care.
A provider-specific detector could become powerful evidence that a particular AI system participated in producing text.
It should not automatically become proof of who authored that text.
Nor does watermark detection itself establish whether AI use breached a contract, academic rule, employment policy or professional obligation.
Those are separate legal questions.
Anthropic's own description makes this limitation clear: detection establishes probable involvement, not authorship or responsibility.
The distinction may become one of the most important consequences of the EU's new transparency regime.
Businesses should distinguish between three questions.
For professional publishers, law firms and other regulated businesses, that final point is particularly important.
Using AI as a drafting tool and blindly publishing AI output are not legally equivalent activities.
YARD Law advises businesses developing and deploying AI systems on the EU Artificial Intelligence Act, including Article 50 transparency requirements, generative AI policies, AI-generated content, GDPR, contractual allocation of AI risk and internal governance.
For businesses publishing AI-assisted content, we can assess whether disclosure obligations apply and design review and documentation procedures appropriate to the organisation's use of generative AI.
Tell us how your business uses generative AI - drafting, translation, customer-facing content or an AI feature in your own product - and we will set out which Article 50 obligations fall on you, and what review and documentation you need to rely on the editorial-responsibility exception.
This article is for general information and does not constitute individual legal advice. Prepared by the legal team at YARD Law Co., a law firm based in Sofia, Bulgaria, working across technology and IT law, data protection and commercial law.
Questions about the EU AI Act and your use of generative AI? Get in touch.
Send an enquiry