How Claude's Text Watermarking Works

Future Claude models will embed a watermark in generated text to help determine if the text was AI-generated, complying with the EU AI Act. The method, based on SynthID-Text, does not affect output quality or readability.

How it works

LLMs pick words one at a time, choosing among candidates. For low-stakes choices (e.g., “overcast” vs “grey”), the selection is random. Watermarking uses a key plus preceding words to seed that randomness, creating a detectable pattern. Detectors with the key can check if the word sequence matches the key-based pattern.

Impact on output

Anthropic reports no quality change, and Google DeepMind's SynthID-Text paper found no statistically significant differences in user ratings. Human raters saw no quality difference. No hidden characters, no added tokens, no extra cost.

Compliance

Implemented to meet EU regulations requiring AI providers to mark AI-generated content. Other major providers signed the same Code of Practice.