If only the vendor can read the watermark, who does AI-provenance detection actually serve?
Anthropic now stamps an invisible mark into every sentence Claude writes, but only Anthropic holds the key, and a hit proves a machine was involved — not who did the work. Does that make provenance a public good, or does it hand a handful of AI companies quiet authority over accusations they alone can adjudicate?
Commentaires (1)
In the August 18 episode of Minds, Bodies, and Terawatts, we dug into Anthropic’s August 14 explainer on watermarking all Claude output worldwide to satisfy the EU AI Act’s Article 50 rules. The uncomfortable part isn’t that the mark exists — it’s that it’s weakest exactly where someone stands to be accused: short samples, factual writing, and code. Research from Kirchenbauer’s team at Maryland suggests you need something like 800 words before the call is reliable, and a cover letter is 200. So we may be building an evidence system that can’t clear the innocent but will still get cited against them. Give the episode a listen and tell us where you land — should provenance keys be vendor-held, publicly auditable, or not exist at all?
Related reading on unscarcity.ai:
Envie d'aller plus loin ?
Obtenez le plan complet dans <em>L'ère de la post-pénurie : Repenser la société à l'ère des machines</em>
