Frontier models can unmask anonymous authors from a single text, UMD paper finds
rohanpaul_ai · x · 2026-11-22
A University of Maryland paper shows frontier models, given a single text and a simple zero-shot prompt (no web, tools, or candidate lists), can determine whether text was AI- or human-written, name the lab behind AI text, and often identify the specific human author.
Key findings:
- Claude Fable 5.2 named the reviewer of 56% of 398 private peer reviews and the account behind 69% of comments from active Redditors (>500 posts/year)
- Opus 5.5 identified the correct lab for 98% of texts from Anthropic, OpenAI, and Google models
- They match trained detectors like Pangram while explaining predictions in natural language
Accuracy tracks the author's exposure in pretraining data and reasoning effort; simply asking a model how familiar it is with someone predicts identification success. The authors warn anonymity for anyone who has published enough online is far weaker than assumed, with limited defenses—blind review and pseudonymous forums should treat author anonymity as broken.
More from Safety
- Google Built a Sealed Internet Replica to Test Browser Agents, 1,000 Attack Tests a Day — shashib · 2026-11-22
- Fields medallist Tim Gowers speaks at London March Against AI Extinction — birchlse · 2026-11-22
- Signal restricts mobile screenshots as it confronts 'existential threat' of AI surveillance — TechCrunch AI · 2026-11-22
- If AI smashes through every bottleneck, unchecked power could arrive in 6-18 months — NathanpmYoung · 2026-11-22
- Right to Compute manifesto pays off: Montana SB212 signed into law — nptacek · 2026-11-22
- NYT: AI Is Exposing Decades of Vulnerable Code as Tech Outages Pile Up — chrismattmann · 2026-11-22