Anthropic is putting a machine-readable trail into Claude’s output, aiming to make AI-written material easier to spot. The catch is that the trail may identify an AI touch without resolving the more combustible question: who, if anyone, actually wrote the work.

The shift follows the EU AI Act’s Transparency Code, which took effect on August 2 and requires providers to make AI-generated or edited content detectable by other systems. Anthropic says Claude models released after that date will carry markings from launch, while support for older models remains in development.¹

For text, the company describes an “imperceptible watermark” embedded without changing meaning, quality or readability. It travels with copied-and-pasted text and “may persist through some editing”; for supported files, Anthropic plans to use the C2PA provenance standard. The marks will apply globally, including through Claude’s API and cloud partners, rather than only in its consumer chatbot.²

That could hand schools, publishers and platforms a sharper clue in disputes over AI-assisted work. Publishing has already seen books pulled or questioned amid allegations of undisclosed AI use, even as authors have denied generating the work themselves.³ Anthropic also says it intends to provide third parties with detection tools.

But the company’s own limits temper the promise. Heavy editing, paraphrasing, translation or mixing Claude material with other writing can make a watermark undetectable. And a detected mark does not establish that Claude authored a passage: proofreading or translation through the model may leave one behind.⁴

That distinction matters most outside classrooms and novels. One commentator argues that workers using AI for emails, meeting summaries, internal documents and Slack messages may be willing to be “caught Claude-handed,” because those tasks do not carry the same authorship expectations as academic or literary work.⁵ Watermarks may make AI use more visible; they are unlikely to settle the norms around it.