Study Reveals SynthID Watermark Easily Detectable and Scrubable
basedjensen · x · 2026-08-26
SRI Lab published an in-depth evaluation of Google DeepMind's SynthID-Text watermarking. Key findings include: (1) Watermark presence is easily detectable via black-box queries; (2) It is more resistant to spoofing than SOTA schemes; (3) Spoofing attempts leave discoverable traces; (4) It is easier to scrub than other SOTA schemes, even for naive adversaries.
More from Safety
- 10-25% of MATS Fellows Are Not 'Safety-Pilled', Estimate Suggests — AaronBergman18 · 2026-08-26
- Misanthrope: A Privacy-Preserving Keypoint Detector that avoids detecting people — zhenjun_zhao · 2026-08-26
- Circumventing AI watermarks directly in-prompt with PRNGs — JulianHabekost · 2026-08-26
- User documents persistent discrepancy between Gemini history and Google Activity records — Aggressive-Arm-1182 · 2026-08-26
- Inside the Relay Market Powering Token Resellers and Fraud — TMWNN · 2026-08-26
- Should I Worry About Cheap Models Training on My Data? — AkindaGood_programer · 2026-08-26