Anthropic’s push to make Claude text traceable promises more transparency—but it may also turn ordinary workplace polishing into an AI disclosure people never intended to make.

The shift follows the EU AI Act’s transparency obligations, which took effect on August 2. Anthropic says new Claude models will embed machine-readable watermarks in generated text and attach signed C2PA provenance metadata to supported files; the marks are meant to work worldwide, across Claude’s consumer products, API and cloud access.¹

The company says the text signal is “imperceptible” and travels when material is copied and pasted, potentially surviving some edits.² Existing products have a four-month grace period, however, so Anthropic is still building support for older models rather than flipping the system on immediately.³

That timing matters because the watermark is not an authorship test. A press release drafted by a human could be marked after Claude merely cleaned up its language, translated it or reformatted it.⁴ Conversely, Anthropic warns that extensive rewriting, mixing text or shortening it can make the marker disappear—leaving platforms with both false suspicions and false reassurance.⁵

For schools, publishers and platforms, the technology offers another lead in disputes over undisclosed AI writing. But it also sharpens a cultural divide. Critics fear a “digital tattoo” for students, journalists and writers who use Claude for limited assistance, while supporters argue the only reason to resist detection is deception.⁶

The more mundane reality may be less dramatic: much AI-assisted writing happens in internal emails, meeting summaries and work documents, where strict literary-style authorship norms rarely apply. The harder question is whether employees will accept disclosure as the price of convenience—or whether watermarking simply starts a new arms race between generators, editors and detectors.⁷