Claude now watermarks its AI-generated text

Anthropic has started embedding a watermark directly into text generated by Claude, effective for every model released since August 2, 2026, across the API, Claude.ai, Claude Code, Cowork, and Claude Tag. Files get a separate treatment: image and document outputs carry C2PA metadata, the open provenance standard other AI vendors already use. Anthropic says the text watermark is built into the model itself and can survive copying and pasting, though it may not survive heavier editing.

The move complies with the EU AI Act’s Transparency Code, which requires AI companies to mark generated or edited content so other systems can detect it. Anthropic says support for watermarking older, already-released models is coming later. TechCrunch reports Anthropic has not yet clarified exactly how much rewriting is needed to strip the mark.

Update, August 16: how the watermark actually works

Anthropic has now published the mechanism. Claude generates text one token at a time, normally picking among several statistically similar next-word candidates at random. The watermark replaces that randomness with a pattern derived from a secret key plus the preceding text, so the resulting sequence carries a detectable statistical signature over a long enough passage, the same underlying approach as Google DeepMind’s SynthID-Text. Anthropic says the watermark adds no extra tokens and has no effect on output quality, cost, or speed, and that anyone holding the secret key can test whether a passage is statistically consistent with keyed Claude output. It cannot confirm a human wrote something, cannot identify text from another AI system, and cannot distinguish original Claude output from text Claude heavily edited.

Update, August 19: removal tools already circulating

Wired reports coders started building removal tools within days of launch. A developer named Meyer built one that uses a separate, non-watermarking LLM to rewrite Claude output with synonym swaps and light reorganization; the code has drawn more than 100 contributors and 20,000 bookmarks on X. Software engineer Erik Hughes built a similar tool in 15 minutes that strips invisible and look-alike characters, reorders sentences, and swaps words. Oxford visiting fellow Leon Chlon says condensing or translating Claude’s output and translating it back also breaks the watermark.

None of the tools has been verified against Anthropic’s own detector, since Anthropic has not yet released one publicly; developers are working from their understanding of the underlying SynthID-style approach, not confirmed test results. An Anthropic spokesperson told Wired:

“We’re adding marking to Claude’s output to comply with the EU AI Act, and other labs are taking similar steps. It’s hard to identify AI-generated text, and this gives people better tools for identification.”

The company says it’s still building its own detection tool and continuing to refine the watermarking system. The EU AI Act rules bar model providers from marketing circumvention tools, but there’s no legal restriction on independent developers building or sharing them.

What it means for you

If your workflow runs Claude output through further editing before it ships, whether that’s a content pipeline, a support macro, or a code comment, treat the watermark as present by default rather than assuming a light copy-edit removes it. This lands the same week Anthropic narrowed Fable 5’s safety false positives, part of a pattern of Anthropic making previously invisible model behavior visible and auditable rather than leaving it opaque. Anyone relying on undetectable AI-generated text for compliance reasons in the EU should test their actual output against the C2PA and text-watermark checks now, not after an audit asks.