Anthropic explains how Claude's new watermarks work

Following its August 11 announcement, Anthropic has now published technical details on how the invisible watermarks in Claude-generated text work. The approach follows the SynthID-Text method that Google DeepMind introduced in 2024: when phrasings are interchangeable — say, “overcast” versus “grey” — the model makes systematic choices that together form a statistically detectable pattern.

According to Anthropic, the watermark does not affect output quality. For verification, the company plans its own detection API — a fundamentally different approach from classic “AI detectors,” which merely hunt for linguistic tells and are notoriously unreliable.

Limits remain: light edits do not fully remove the pattern, but a complete rewrite in which every word is replaced does. For code, the method barely applies, since working code leaves little linguistic slack — comments may carry watermarks, the executable code itself does not.

A key driver is the EU AI Act: its transparency rules require that AI-generated content be identifiable as such. Anthropic thereby positions itself as the first major US provider to roll out text watermarking at scale.

Reactions are mixed: in forums, some subscribers announced cancellations, worried that watermarks could expose unauthorized use at work or in exams. The debate over detectability versus privacy is likely to occupy the industry for some time.