TechCrunch reports that Anthropic's new watermarking system for Claude output has produced a wave of social-media complaints — from users who object that the marks will reveal their use of the model to employers and instructors.
Which is, of course, the feature working as specified. Output provenance marking is only useful if it is legible to someone other than the person who generated the text, and the complaint is a fairly precise description of the threat model it was built for.
The framing shifted in the telling: TechCrunch's URL slug still reads "catch them cheating at their jobs, classes" while the published headline settles for the softer "catch them using it." The Hacker News thread drew 60 points and 86 comments.
Details on how the watermark is embedded, what it covers, and how robust it is to paraphrasing are not in the coverage — worth waiting for before drawing conclusions about how much it will actually catch.
Sources: TechCrunch, HN discussion