Anthropic says text watermarking scheme relies on inconsequential words

August 15, 2026

Anthropic’s watermark influences low-stakes word choices using Google DeepMind’s SynthID-Text technique. A different source of randomness creates a detectable statistical signature.

Light editing may not remove the watermark; complete rewrites will. Anthropic claims no impact on quality, speed, or pricing. Detection API planned; robustness against evasion remains debated.