Anthropic’s Claude watermark goes further than rivals

Anthropic’s new watermark for AI-generated text in Claude is prompting debate over whether the company is taking a broader approach than required under European rules. Under the European Union’s Article 50(2), AI companies do not have to mark outputs when systems are used for standard editing or when they do not substantially change a user’s original words or meaning.

Claude’s watermark guides word choices to create a detectable pattern that can remain when text is copied and pasted. That means it can appear in translations, summaries, and more extensively edited text, raising concerns that readers may assume Claude wrote a document even when it was used only for support. Anthropic said lighter edits are less likely to trigger the watermark because there is little for the signal to attach to, and said it will “continue to evaluate different approaches.”

The EU’s watermarking requirements apply to new models launched in Europe on or after August 2, while older models have a grace period until December. Google already uses SynthID for AI-generated text, images, audio, and video, while OpenAI plans to expand labels to text. Meta has focused on images, and xAI did not respond to a request for comment.