Anthropic is adding a Claude text watermark so people can later check whether Claude likely touched a passage. The company’s August 14 explainer says the mark is invisible, adds no hidden characters, and does not change price or speed. The company is doing it to meet the EU AI Act’s marking rules, and it is turning the feature on worldwide because it cannot yet scope it by region.
How the Claude text watermark works
Claude already picks the next word from a list of good options. On low-stakes choices, it used to settle the tie with a random number. Specifically, the watermark swaps that randomness for a secret key plus a few prior words. A reader still sees ordinary prose. Someone with the key can later score how likely the sequence is Claude’s.
Anthropic says the method is a version of Google DeepMind’s SynthID-Text, published in Nature in 2024. DeepMind’s Gemini tests found no statistically significant quality drop. Anthropic also says its own testers saw no hit to content, creativity, or readability. Meanwhile, the mark carries no user, company, or chat identity.
What it can prove, and what it cannot
“A watermark can only determine that Claude was likely involved with the content at some point.” It cannot tell “Claude wrote this” from “Claude heavily edited this.” Translations get marked, because Claude chooses every word. Light edits probably keep the mark. A full rewrite that replaces every word will not. Short passages and tightly factual lines give the watermark almost nothing to sit on.
However, code is thinner still. Where an exact token is required, the watermark does not nudge the model. Comments can carry a mark. The generated program usually will not. Besides text, supported files such as PNG, JPG, and SVG get a signed C2PA content credential in metadata. That label is not a watermark. It only says Claude processed the file.
Then there is timing. Anthropic’s help article says models launched on or after August 2, 2026 get marking at launch, across Claude, the API, Claude Code, Cowork, Tag, and the big cloud partners. Older models sit in a transition period. The company says those watermarks will roll out over the coming months. A detection API is coming; it is not live yet.
As a result, this is the opposite of Google’s visible Gemini badge. TMM already covered how Gemini now lets users hide those on-screen marks. Claude’s mark lives inside the words themselves. Also, it will sit on every supported surface, including Claude in Chrome, not just the chat box. The company still treats the result as a likelihood test, not a courtroom stamp of authorship.
