Anthropic watermarks Claude text with SynthID

oreillyradar.substack.com16 min read

Major AI companies like Anthropic and Google now watermark text generated by their models to mark machine-written content, but they do not allow the public to check for these watermarks. The author explains how text watermarking mechanisms work, detailing how models output token probabilities and how runners modify the selection process using secret keys. By building a practical web toy based on the SynthID method, readers can understand how these watermarks are embedded and detected, while also seeing why companies keep their keys private to prevent tampering. Open article →