Researchers Find a Loophole in Claude's Watermark

Claude AI is made by Anthropic. The company embeds an invisible watermark in the text its models generate. This is meant to comply with the EU AI Act, whose requirements take effect in December. But on two of its older models, Claude Opus 4.6 and Claude Sonnet 4.6, that watermark can be removed. All it takes is changing one parameter used during generation: temperature, top_p, or top_k. These control how random the model's word choices are. According to Anthropic's own documentation, these settings are supposed to stay at their defaults for the watermark check to work reliably.

So keep this in mind: watermarks in AI-generated text only hold up if none of the default parameters have been changed. Changing just one is enough to make the watermark disappear. The watermark exists mainly to meet regulatory requirements. Adjust the temperature during generation, and you can strip it out cleanly, with no trace left behind. So I would not lean on a watermark like this too much. It is quite possible for text written by AI to carry no sign at all that it was "marked as AI."