Anthropic’s push to make Claude-generated material traceable promises a win for transparency—but it may also turn ordinary workplace AI assistance into a visible signal that users never meant to disclose.

The change follows transparency obligations under the EU AI Act that took effect on August 2. Anthropic says new Claude models released after that date will carry machine-readable markings from launch, while older models are due to be brought into compliance during the grace period. The system will apply worldwide, across Claude’s consumer products, API and cloud-provider access.¹

For generated files, Anthropic plans to use C2PA provenance metadata. For text, it is embedding an “imperceptible watermark” that does not alter readability and can follow copy when it is pasted elsewhere. “Watermarking will be applied at the model level,” Anthropic said, meaning it will appear regardless of which Claude surface produced the text.²

That breadth is precisely where the policy gets thorny. A communications team that asks Claude to polish, translate or format a human-written release could leave behind an AI marker, even when Claude did not write the underlying copy. The same technology may help schools, publishers and platforms scrutinize work passed off as wholly human—but a detected mark will not, by itself, prove AI authorship.³

Anthropic also acknowledges the reverse problem: heavy rewriting, translation, mixing text with other material or simply using short passages can make a watermark difficult to detect. That leaves the industry with a familiar arms-race dynamic, while platforms from Substack to LinkedIn experiment with their own ways to identify AI-made material.¹

Some observers argue the social consequences may be narrower than the alarm suggests. AI-written academic work and novels remain especially fraught, but much everyday use involves meeting notes, internal documents, emails and summaries—tasks where users may be more willing to be “caught” using a tool.⁴ Claude’s labels could therefore make deception harder while accelerating a more uncomfortable question: when does assistance become authorship?