Anthropic Confirms Claude Watermark Uses SynthID: No Hidden Characters or User Tracking
08/15 04:25
According to monitoring by Beating, Anthropic has finally clarified the mechanism behind the Claude text watermark. Previously, the company only mentioned the addition of an invisible watermark, but it has now confirmed the use of Google DeepMind's SynthID-Text technology. This method does not embed zero-width characters or hidden codes into the text. When Claude encounters multiple plausible next words, it slightly alters the selection probability using a key. This change accumulates throughout the text, ultimately forming a detectable statistical signal. The watermark does not increase tokens, has minimal impact on generation speed, and does not change pricing. Additionally, the watermark cannot be traced back to users. The detection results can only determine whether 'Claude has participated in generating this text,' without revealing who, which company, or which conversation produced it. It only allows Claude to modify grammar and punctuation; if the changes are too minor, detection may not be possible. The watermark signals left by the code are also weaker, but Claude will still carry the watermark when translating the entire text. Since the watermark was implemented, some users have canceled their Claude subscriptions, but Anthropic claims there has not been an overall increase in cancellations. The company will also open a watermark detection API, allowing users and third parties to check the text themselves.
全部评论0最新最热