Stanford paper proposes differentially private watermarks for LLMs
ChengleiSi · x · 2026-11-06
Anthropic and OpenAI will watermark their responses, but critics argue this violates user privacy and incentivizes paraphrasing away the watermark — a lose for both sides.
A Stanford paper by Elise L. Han and Tatsunori Hashimoto, "Reconciling Privacy with Provenance for Language Model Watermarks," reframes the incentive structure: shift the AI provenance goal from individual tracking to estimating aggregate AI usage. The authors develop locally differentially private watermarks that provably limit the evidence any individual text can contribute to detection, while still enabling accurate estimation of the population prevalence of watermarked texts.
More from Safety
- METR report: Anthropic is speeding up, not pacing, and redacted the key parts — asusarla · 2026-11-06
- Polymarket odds: only ~9% chance US enacts an AI safety bill by year-end — Polymarket · 2026-11-06
- Man sues OpenAI, alleging ChatGPT helped his wife plot to poison and kill him — Polymarket · 2026-11-06
- Alignment researchers debate threat models beyond scheming — MariusHobbhahn · 2026-11-06
- Sentinel Bio's Joshua Monrad on DNA screening, uplift studies and agent monitoring — The Cognitive Revolution · 2026-11-06
- Sentinel Bio's Joshua Monrad on AI Biorisk: DNA Screening, Uplift Studies, Agent Monitoring — The Cognitive Revolution · 2026-11-06