AI Text Watermarks and the EU AI Act: Who Is Liable?

Since 2 August 2026, every Claude model has embedded an invisible statistical watermark in the text it generates. Plenty of business owners read this as a new compliance burden landing on them. It is the opposite. This duty sits with the company that builds the model, not with the business that uses it to draft a quote, a product sheet, or an internal report. This applies regardless of where your business is based, as long as you operate in the EU or sell into it.

What did Anthropic actually change on 2 August 2026?

Anthropic announced, in coverage that broke on 11 August 2026, that every Claude model released from 2 August 2026 onward now embeds an invisible statistical watermark directly in generated text (Fortune, 11 Aug 2026). The rollout spans the whole product line: the API, claude.ai, Claude Code, Claude Cowork, Claude Tag, and Claude models accessed through AWS, Google Cloud, and Microsoft Foundry.

The mark is imperceptible when reading the text. It does not change the meaning or quality of the output. It survives copy-paste from one document to another, and can survive some later edits (Forbes, 13 Aug 2026). Anthropic is being described as the first major lab to ship a text watermark at this scale across its entire product line, rather than on a single entry point.

The launch has not gone unnoticed: some uses of the watermark to detect AI-written text in school or work settings have drawn a sharp reaction from users (TechCrunch, 12 Aug 2026). That friction does not change the regulatory reading below. It just explains why the topic is getting attention right now.

Why now? What is the regulatory trigger?

The trigger is regulatory. Article 50 of the EU AI Act, in force since 2 August 2026 for newly placed systems, requires providers of generative AI systems to embed machine-readable marking in their output, so platforms and authorities can identify generated content (Regulation (EU) 2024/1689, Article 50, EUR-Lex). Anthropic's rollout date lines up exactly with that application date.

One point needs stating plainly: the Digital Omnibus, in force since 27 July 2026, pushes some high-risk obligations to December 2027 and August 2028, but it does not touch Article 50. Nothing in that simplification package delays the watermarking duty. If you have heard that the watermark requirement was softened or postponed, that is not accurate.

Who actually carries the duty: the provider or the business?

This is the confusion worth clearing up. Article 50(2) of the EU AI Act places the machine-readable marking duty on the provider of the generative system, the entity that builds and places the model on the market. A business drafting content with Claude, ChatGPT, or any comparable tool is a deployer. It has no watermark to build. The provider has already put it there, upstream of any use the business makes of the tool.

This provider-versus-deployer split runs through much of the AI Act, as covered in more depth in the EU AI Act for SMEs. Here it has a direct payoff: the text-watermarking story does not add a task to your compliance list. It removes a worry that was never yours to carry.

Does a business still need to disclose that content is AI-generated?

A second, separate duty exists: disclosure under Article 50(4), second subparagraph, for published text that informs the public on matters of public interest. The useful part for a business is that this disclosure duty lifts once the content has gone through human review with someone taking editorial responsibility for publishing it.

In practice, if a text drafted with a model's help gets reviewed, corrected, and signed off by a person who owns the decision to publish it, blanket disclosure is not required. The actionable takeaway follows directly: do not stamp an "AI-generated" label on every page. It is not required in that case, and it tends to undercut the credibility of content that was, in fact, reviewed and approved by a human. What is actually worth having is a written usage policy: who reviews, who signs off, what is off-limits.

Is the text watermark the same thing as the deepfake rule?

No, and the distinction matters. Article 50(4), first subparagraph, covers a separate and stricter regime: deepfakes, meaning content that imitates the appearance, voice, or statements of a real, identifiable person. A generic synthetic voice used for a business audio clip, for instance, does not clone any specific individual and does not fall under that tougher regime.

It helps to keep three layers apart within a single article of the regulation: the invisible technical mark (provider's job, already handled), disclosure of published content (deployer's job, conditional on human review), and deepfakes of a real person (a separate, stricter regime). Mixing the three leads either to unnecessary anxiety or to carelessness in the wrong place.

What should a business actually take from this?

The Claude watermark does not create any new duty for a business using the tool. What matters is what already mattered: knowing who reviews content before it goes out, writing that rule down, and reserving an explicit "AI-generated" label for cases where no human editorial review took place. The broader compliance picture, risk tiers included, is covered in the EU AI Act for SMEs.