Two Minute Papers · 2026-09-15 · notable
Two Minute Papers — 'Claude Is Now Leaving Invisible Fingerprints In Its Text'
Two Minute Papers posted 'Claude Is Now Leaving Invisible Fingerprints In Its Text' on 15 September 2026. The subject named in the title is Anthropic's text watermark, which bends Claude's word choices so its output carries a detectable pattern.
Two Minute Papers takes on Claude's invisible text watermark in an episode posted on 15 September 2026.
What is it?
'Claude Is Now Leaving Invisible Fingerprints In Its Text' went up on the Two Minute Papers channel on 15 September 2026. The thing named in the title is the text watermark Anthropic added to Claude — an imperceptible signal carried by the words themselves, which Anthropic described in a post first published on 14 August 2026 and updated on 1 September 2026.
How does it work?
Anthropic's watermark rides on word choice, not on hidden characters, so it survives copy and paste. Wherever the model has several equally good next words, the usual random number generator is swapped for a function of a secret key and the few words already written, so the choices leave a repeating statistical pattern once a passage is long enough. Anthropic says the method is a version of SynthID-Text, published by Google DeepMind in Nature in 2024, in a line of work going back to a 2022 proposal by Scott Aaronson.
Why does it matter?
Detection is the part most people actually want, and it is not open to everyone: Anthropic's detection API is in private preview for eligible organisations such as regulators, law enforcement, media, fact-checkers, researchers and educational bodies, as required under EU law. Watermarking also applies to future Claude models first, with older ones being added over the coming months. A channel with a large general audience walking through the mechanism reaches readers and writers who will never open Anthropic's technical post.
Who is it for?
writers and developers publishing Claude output