Anthropic Confirms Claude Watermark Using SynthID: No Steganography, No User Tracking

According to Dynamics Beating monitoring, Anthropic has finally explained the mechanism of Claude's text watermark. Previously, the official statement only mentioned the addition of an invisible watermark, but it is now confirmed that they are using Google DeepMind's SynthID-Text technology.

There will be no zero-width characters or hidden code inserted into the text. Whenever Claude encounters the next word in the text that makes sense, a slight change in the selection probability will be made using a key. These changes accumulate throughout the entire text, forming a statistical signal detectable by machines. The watermark does not increase the token count, almost does not affect the generation speed, and the price remains unchanged.

The watermark also does not allow user de-anonymization. The detection results can only determine whether "Claude has participated in this text," without revealing the identity of the user, company, or the specific chat segment generated. If Claude only makes minor changes to the grammar and punctuation, the watermark may not be detectable; the watermark signal that code can leave behind is also weaker. However, a full-text translation by Claude will still contain the watermark.

After the watermark was launched, some users canceled their Claude subscriptions due to this change. However, Anthropic stated that they have not seen a significant increase in overall unsubscriptions. The company plans to release a watermark detection API next, allowing users and third parties to check the text themselves.