Anthropic Watermarks AI Text

San FranciscoAnthropic will swap word choices to mark Claude’s text.

Detection requires Anthropic’s own key, so no outsider can independently verify a watermark claim.

The system changes which word a low-probability random choice lands on, like “overcast” versus “grey,” leaving no visible trace.

Anthropic has not said when watermarking ships or whether output quality was tested against it.

Sources: 404 Media