Anthropic details how watermarking will work in Claude's text

Anthropic published a blog post on Friday explaining in more detail how it plans to watermark text generated by its Claude chatbot. The announcement follows the company's earlier statement this week that it would introduce watermarking to comply with the EU AI Act's Transparency Code, which requires AI developers to use systems enabling identification of AI-generated content. The move sparked debate among users, with some on Reddit calling it an attack on loyal customers and others defending it as a matter of honesty. According to Business Insider, a number of users on X claimed to have cancelled their Claude subscriptions over the change.

How the watermark works

Anthropic explained that when Claude faces "low-stakes choices" — such as deciding between the words "overcast" and "grey" to describe weather — it can create a subtle pattern in its word choices that is invisible to readers but detectable by anyone holding a special key. The company stressed that watermarking does not affect the quality of Claude's output and that a watermarked response looks identical to an unwatermarked one. Anthropic said it will use the SynthID-Text approach, originally outlined by Google DeepMind in 2024, and plans to release a detection API. This differs from AI-detection tools offered by companies like Pangram, which instead look for stylistic "tells" common in AI-generated writing.

Editing and code

According to Anthropic, light editing of watermarked text likely won't remove the watermark entirely, while a complete rewrite replacing every word will — though in that case, it becomes debatable whether the text can still be called AI-generated. If Claude only proofreads human-written text, whether the watermark remains detectable depends on the length of the text and the extent of the edits. Code will carry a weaker watermark than other text, since the model must produce functioning code with limited freedom to choose between alternatives, though it can still appear in areas like code comments. Anthropic also noted that other major AI developers who signed the same Code of Practice will roll out their own watermarking systems as well.