Text Gets Skewed Word Choices, Files Get a Signature
The text watermark works by statistically biasing Claude's word choices according to a key Anthropic holds: any single choice looks unremarkable, but across enough text the pattern becomes detectable, and it travels with the text when it is copied and pasted elsewhere. Anthropic says this does not change the meaning, quality, or readability of a response. On the file side it uses the open C2PA standard, attaching signed provenance metadata to supported types such as `.svg`, `.png`, and `.jpg` — a label that signals the file was processed by Claude and lets you detect whether it has been tampered with since. Because the marking sits at the model layer, it follows the output through every surface: the Claude Platform API, the web and desktop apps, Claude Code, Claude Cowork, Claude Tag, and versions accessed via AWS, Google Cloud, and Microsoft Foundry. Anthropic says it is also working to add marking support to models released before that date.
Compliance Is the Driver, and the Limits Are Spelled Out
The trigger is Article 50(2) of the EU AI Act and the transparency code that took effect on August 2, requiring providers of systems that generate synthetic text, audio, image, or video to mark outputs in a machine-readable format detectable as artificially generated. Anthropic chose to apply this worldwide rather than only for EU users. The company lists the limits itself: a detected mark indicates the content *may* have been processed by Claude, not that Claude produced the ideas or the writing — proofreading, translating, or summarizing a passage can all leave a trace. In the other direction, heavy editing, paraphrasing, translation, short passages, stripped metadata, and unsupported platforms can all defeat detection, so an absent mark does not prove a human wrote it.
The Backlash Is About the Missing Pieces
The objection centers on false accusations. Anthropic says it will support third-party detection as the code requires, with details in forthcoming documentation, but no detection tool, accuracy threshold, or dispute process is public yet, and developers cannot wire detection into their own compliance pipelines. The result: a document a person wrote and merely had Claude polish can be flagged as "AI-generated," with no checkable evidence available to the author, and some users have said publicly that they cancelled subscriptions over it. Where output gets screened — submissions, coursework, client deliverables — the safer move for now is keeping your own record of the drafting process.
via: Claude Help Center, "How Claude marks AI-generated content", TechCrunch report, Forbes report; verified 2026-08-15