SynthID Watermarking in Claude Costs Three Points of Code Correctness on One Model, but Detection Is Near Chance

arXiv 2609.09604·high signalXBlueskyLinkedInCopy link

arXiv 2609.09604 (2026-09-09) responds to Anthropic's disclosure, days after EU AI Act Article 50 took effect on 2026-08-02, that every Claude model released after that date embeds a SynthID-Text watermark by default with no opt-out. Because no public tool can test the deployed systems, the authors evaluated the open-source SynthID-Text implementation on two open-weight models: on prose the measured effect does not exceed changing the sampling seed, and on code the cost is three points of correctness on one model and below measurement on the other, while detection stays near chance. Their argument is that the unverifiability, not the watermark, is the governance failure.