A five-hour watermark remover, 14,000 stars, and no detector to test either side

The useful split inside the repository is between what a stranger can check and what nobody can. Stripping invisible Unicode characters and C2PA, EXIF, XMP and document metadata is verifiable on the spot: run the tool, open a metadata reader, see whether the fields are gone ^[9]. The statistical text watermark is different, because the signal lives in word choices, and Meyer's own documentation calls the rewriting path best-effort and says it cannot certify that a vendor detector will fail ^[11]. The loud half of this story is the half with no scoreboard.

Anthropic's half is unscored in exactly the same way. The company says the text mark travels through copy and paste and may survive some editing ^[6], while also conceding that its marks cannot establish authorship and may disappear after heavy editing ^[7]. Both statements describe the behaviour of a detector that outside researchers cannot run, because the detection mechanism and full technical guidance have not been published ^[8]. Durability claim and removal claim rest on the same missing instrument.

That is what makes the marking a deterrent rather than a control. A signed C2PA manifest is a positive assertion attached to a PNG, JPEG or SVG ^[5]; delete it and you do not get a file flagged as tampered with, you get a file that looks like every other unmarked file ^[9]. Provenance metadata raises the cost of casual laundering and creates something a platform can point at when it is present. It cannot answer the question people actually ask, which is whether unlabelled text came from a model, and Anthropic has already said its marks do not settle authorship ^[7].

The effort asymmetry is the number worth holding onto. Anthropic's marking programme spans Claude, Claude Code, Claude Cowork, Claude Tag, its API and supported cloud platforms, with models launched in the EU on or after August 2, 2026 supporting machine-readable marking from launch ^[4]. Against that, one developer with more than 20 years in technology and prior experience running open-source models produced a first version in roughly five hours ^[2][15]. The 14,000-plus stars work out to about 2,800 per hour of build time ^[18], which measures curiosity and nothing about efficacy; Meyer himself says the attention caught him unprepared ^[14], that the tool is still difficult to use, and that it will need months of adjustment as Anthropic ships or changes detectors ^[13]. The repository warns that removal is not guaranteed ^[17].

Untestability is also a business model. Unmarker is a pre-launch commercial service advertising Claude-focused scoring, rewriting, meaning checks and planned API access, and Unmarkr claims to remove visible marks, hidden AI tags, SynthID and C2PA metadata from images ^[16]. A paid service that says it defeats a statistical watermark is selling a claim its customers cannot check and reviewers cannot check either, for as long as the only working detector sits inside the lab that made the mark ^[8]. Meyer's stated purpose, educational use and people working on their own content ^[12], is a README sentence, not a constraint on who clones the code.