Anthropic will embed watermarks in the text its models generate, including Claude, to satisfy European transparency rules. TechCrunch reported the change on August 11, citing an updated Anthropic support page. The company presents it as compliance. That framing is accurate, and it is the most useful lens for reading the rest.
The rule driving it took effect on August 2. Under the transparency code attached to the EU AI Act, anything a model generates or edits has to carry a mark that other systems can read. Anthropic said every model shipped after that date automatically marks generated text and generated files. Files use C2PA, the open provenance standard. Older models will get the capability too, according to the support page, which as quoted by TechCrunch does not attach a date to that.
Scope, exactly as stated. Marking is applied at the model level, so it is present regardless of which Claude surface produced the text. The named products are Claude itself, Claude Code, Claude Tag, Claude Cowork, and the Claude platform API. The support page says the mark is part of the text itself, so it moves with a copy and paste, and “may persist through some editing.”
That final clause is doing an enormous amount of work.
Text watermarking is a much harder problem than image watermarking, and the difference is not incremental. An image carries millions of pixels of redundancy, which gives a provenance signal room to survive resizing, recompression, and cropping. A paragraph of prose has no comparable slack. Ordinary human editing does not just distort the carrier, it replaces it. A person who rewrites a draft in their own phrasing is not attacking a watermark; they are doing what writers do, and the signal degrades as a side effect. So the interesting number is the one nobody has published: how much rewriting is enough. TechCrunch asked Anthropic to clarify and had not received an answer at publication.
Note who this is built for. The person generating the text gains nothing from being marked. The value accrues to whoever wants to check text later: a university reviewing coursework, a publisher screening submissions, a hiring team reading applications, a platform enforcing policy, a regulator auditing a filing. That asymmetry is the story. This is an institutional detection feature that every individual user now carries, which makes it a compliance artifact first and a product second. Nobody will renew a subscription because of it.
A mark also means nothing on its own. It means something only when a party other than Anthropic can run the check. C2PA gets this right for files, because the standard is open and anyone with a compatible tool can read the manifest. For text, TechCrunch’s report gives no indication that Anthropic has released a detector, described an interface, or said who is permitted to query one. Until that exists, “this text is watermarked” is a claim that only its author can confirm, which is a weak foundation for the academic and editorial decisions people will want to make with it. The report also does not say whether users or API customers can opt out. Treat both as open, not as resolved in your favor.
The pattern extends past Anthropic. TechCrunch noted that Suno said last week it would mark tracks made on its platform after legal challenges, and that Substack partnered with Pangram last month to flag AI-generated content, with chief executive Chris Best naming “Claudefishing” as the behavior. OpenAI, Google, Microsoft, Meta, Synthesia, and Black Forest Labs have all committed to the EU code.
For anyone shipping customer-facing copy through a Claude endpoint, the practical change starts now: assume model output generated after August 2 carries a machine-readable mark you cannot currently inspect, and that your downstream editing may or may not remove it. Before your next release, get two answers in writing from Anthropic: whether an opt-out exists for your tier, and when an independently runnable detector arrives.
Reported by Ivan Mehta for TechCrunch on August 11, 2026.
