Anthropic starts watermarking Claude's text output to meet EU rules

Anthropic has explained the mechanism behind a new watermarking system applied to text generated by its Claude AI models, introduced to satisfy the European Union's recent AI transparency regulations. According to the company, the watermark carries no visible signs, cannot be noticed by readers, and does not rely on inserting hidden characters into the text.

How the system works

Large language models generate text by selecting each word from a list of contextually suitable options, typically choosing somewhat randomly. With watermarking enabled, Claude instead uses a specific key to guide these word choices rather than a random number generator. Anthropic illustrated this with an example using the digits of pi as the key, where each digit determines which ranked word option the model selects next. The approach is adapted from Google DeepMind's SynthID-Text technique, described in a paper published in Nature.

Anthropic says the watermarking process does not reduce output quality, slow generation, or require additional computing resources. The company plans to release an API that uses matching keys to verify whether a given piece of text was produced by Claude.

Limitations

The method has notable constraints. It cannot distinguish whether Claude originally wrote a text or merely edited it — edited and translated text will carry the same watermark. Very short passages or text that Claude only lightly proofread may not retain a detectable watermark. Minor edits to Claude-generated text are unlikely to strip the watermark; removing it reliably would require a complete rewrite.

Regarding code, Anthropic notes that watermarking tends to be weaker because code generation often demands exact, predetermined output with little room for word-choice variation, limiting where a watermark pattern can be embedded. The company will also mark Claude-generated images by adding a cryptographically signed note within the image metadata.

The new system applies to all Claude output worldwide from models released after August 2, since Anthropic lacks a reliable way to limit the feature by region. Older Claude models will gain watermarking capability over the coming months.