Artificial intelligence

Anthropic watermarks Claude AI text to comply with EU law

Published Aug 15, 2026, 6:14 PM2 min readNewUJ Editorial Desk

Anthropic watermarks Claude AI text to comply with EU law
Photo: Logan Voss · Unsplash
0 0
XWhatsAppTelegramLinkedIn

Anthropic said on August 15, 2026 it has begun watermarking text generated by Claude to comply with the European Union’s new AI transparency rules.

The watermark leaves a pattern in the text that is invisible to readers and can be decoded only with a specific key.

Changes apply to all Claude products using models released after August 2, 2026. Anthropic said the watermark does not alter output quality, slow generation, require extra tokens, or make generations more expensive.

Anthropic’s technique adapts Google DeepMind’s SynthID-Text method, described in a paper published in Nature. Normally, large language models choose each word randomly from a list of contextually appropriate options; with watermarking on, Claude uses a key to choose the next word. In an example based on the digits of pi, if the key starts with 2 from the sequence 3.1415926535, the next word is the sixth in the list, then the fifth, the third, and the fifth again.

The company will release an API with keys to decode the watermark and determine whether Claude generated a block of text. It acknowledges limits: the method cannot tell whether Claude wrote the text or merely edited it, so translations and lightly edited text will also be watermarked. If Claude only proofreads and lightly edits a text, or if the text is too short, the watermark may not be detectable. Lightly editing Claude-generated text probably will not remove the watermark; a complete rewrite is needed to be sure.

For code, watermarking is generally weaker than for other text because code often requires exact output. If no choices are available in text generation, watermarking cannot be applied. Anthropic will also watermark images by adding a cryptographically signed note in metadata stating that Claude generated the image.

Anthropic is applying watermarks to all Claude output at launch because it has no reliable way to implement the changes by region. Other AI companies are also required to watermark content under the EU rules. The company plans to add watermarking capability to older Claude models in the months following the announcement.

Sources

Report / request removal

Related

Comments

No comments yet. Be the first.