How Claude’s text watermark works

Future Claude models will generate text containing a watermark to comply with the European Union Artificial Intelligence Act. Anthropic is implementing this change globally to meet the EU requirements that took effect in August 2026. The watermarking method is based on SynthID-Text, an approach published by Google DeepMind in 2024.

The watermarking process works by altering the source of randomness used when Claude selects between equally suitable words. Instead of an arbitrary random number generator, the model uses a specific key and preceding words to determine word choices. This leaves an undetectable pattern in the text that can be identified by anyone with the corresponding key to determine the likelihood that Claude was involved in writing it.

The watermark does not impact the quality, creativity, readability, speed, or cost of Claude outputs. It adds no hidden characters, does not require extra tokens, and carries no identifying information about users, organizations, or chats. The watermark is less effective on short passages, factual texts, computer code, or lightly edited human writing where the model makes few arbitrary word choices. Anthropic will soon offer a watermark detection application programming interface and plans to apply watermarking to older models over the coming months.