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:

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.

Related event: Maryland study: frontier models can unmask anonymous authors from a single text(2 posts)→

Original post →

More from Safety

Safety channel →