Anthropic explains how Claude’s upcoming text watermark will work

Anthropic says future Claude models will include an invisible text watermark designed to estimate whether Claude helped produce a passage. The company is introducing the system alongside other major AI providers as part of compliance with the EU AI Act.

A watermark built into word selection

Claude generates text one token at a time, choosing among plausible next words. The watermark does not add characters, metadata, or extra tokens. Instead, it uses a key to influence the source of randomness behind low-impact word choices. Those choices should remain natural to readers, but the resulting sequence can later be checked for statistical consistency with Claude’s key.

Anthropic says internal testing found no meaningful effect on quality, creativity, readability, or content. The approach is based on Google DeepMind’s SynthID-Text technique, which was evaluated in comparisons of watermarked and ordinary model responses.

For AI builders, the system offers a potential way to assess model involvement without exposing a user’s identity. A successful check would indicate the likelihood that Claude contributed to the writing—not prove that the text was entirely AI-generated, identify a particular person or conversation, or rule out other models.

Detection also has practical limits. Short passages provide too few word choices for strong confidence, while factual statements and light proofreading offer fewer opportunities for the watermark to appear. Heavier Claude-generated rewriting should provide more signal than minor edits to human writing.

Anthropic says the watermark will not add expense and will not be unique to Claude, as providers serving the EU move toward AI-content labeling requirements.

Source: Anthropic News