Anthropic has announced that future Claude models will generate text containing an invisible watermark designed to help determine whether Claude was involved in creating a piece of writing.
The change is being introduced as part of Anthropic’s compliance with the EU AI Act, which requires AI providers serving the European market to mark AI-generated content.
Unlike a traditional watermark, nothing visible will be added to Claude’s responses. There are no hidden characters, tracking codes, extra tokens, or personal identifiers embedded in the text.
Instead, the watermark works through subtle statistical patterns in the words Claude chooses while generating a response.
How the Watermark Works
Large language models generate text by repeatedly selecting the next word or token from several possible choices.
When multiple words would work equally well, Claude’s watermarking system can influence how that choice is made using a secret key. Over a long enough passage, those decisions create a detectable statistical pattern.
Readers should not notice any difference.
Anthropic says the watermark does not meaningfully affect the quality, creativity, readability, speed, or cost of Claude’s responses.
The company is using a version of SynthID-Text, a watermarking approach originally developed by Google DeepMind.
What the Watermark Can — and Cannot — Tell You
A detected watermark can indicate that Claude was likely involved in creating or heavily editing a piece of text.
It cannot prove that Claude wrote the entire document, identify the person who used Claude, reveal an organization or account, or determine whether another AI system created the text.
Detection also becomes less reliable with very short passages, lightly edited human writing, factual material, and computer code because there are fewer opportunities for Claude to make flexible word choices.
Translations generated by Claude, however, can contain the watermark because Claude is selecting the wording throughout the translated text.
Can the Watermark Be Removed?
Yes, to some extent.
Minor editing may leave enough of the watermark intact to remain detectable, while completely rewriting a passage could remove the pattern.
For that reason, watermarking should be viewed as an additional signal of AI involvement rather than absolute proof of authorship.
Why Anthropic Is Making the Change
Anthropic says it is implementing watermarking to comply with transparency requirements under the EU AI Act.
The company signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026 alongside other major AI providers.
Anthropic plans to introduce the watermark globally rather than limiting it to European users.
The company also says a watermark detection API is coming, which will allow users and organizations to check whether text contains evidence of Claude’s watermark.
For images and certain other files created by Claude, Anthropic will use C2PA content credentials instead, adding cryptographically signed information to the file’s metadata indicating that Claude was involved in creating or processing it.
For everyday Claude users, the change should be largely invisible. The text will look and read the same—the difference is that future Claude-generated writing may carry a technical signal that can help identify AI involvement after the fact.
Leave a Reply