Supported Claude models embed a statistical watermark in generated text, Anthropic says. Its documentation now includes model-by-model coverage and access to a detection service in private preview.
The current help page distinguishes text watermarks from Content Credentials: signed C2PA metadata attached to supported generated files. Coverage varies by model, platform and file type, so it is inaccurate to assume every Claude output carries the same marker.
How the text marker works
In its August 14 technical explanation, Anthropic identifies its method as a version of Google DeepMind’s SynthID-Text. It changes the source of randomness used when choosing among plausible words, producing a pattern that can be tested with a key.
Anthropic says this adds no hidden characters or extra tokens and contains no information identifying a user, organisation or chat. Its statements about unchanged output quality are the company’s claims, not an independent finding by Nexus News International.
What detection can establish
A detected mark can indicate that Claude helped generate or process text. It cannot establish who originated every idea, or distinguish original generation from substantial editing.
Anthropic also cautions that short passages and heavily changed text may not produce a reliable signal. A missing watermark is not proof of human authorship.
What has changed since August
Detection is now in private preview for eligible organisations, including certain media, research, education and regulatory users. It is not described as a universally available public checker.
Check the help page’s current support table before drawing conclusions about a particular model. The distinction matters when assessing edited human work: a provenance signal should not be treated as a complete authorship record.
Updated September 25, 2026: Added Anthropic’s technical explanation and current detection availability; corrected the earlier statement that these details were still forthcoming.