arXiv: Opus 5.5 names the lab behind 98% of AI-generated texts from one sample

rohanpaul_ai · x · 2026-11-22

The arXiv version (2611.26216) of the University of Maryland paper adds full author list and abstract details: with a single text and a zero-shot prompt, frontier models answer "who wrote this?" at three levels—AI vs human (Gemini 4 Carbon detects 97% of fully AI-generated texts, matching Pangram 4.1's 97%), which lab (Opus 5.5: 98% for Anthropic/OpenAI/Google texts), and often the exact human author (Fable 5.2: 39% of public Oct–Nov 2026 peer reviews, 78% on a Nature subset; 69% of active Redditors' usernames).

Accuracy depends on pretraining exposure to an author's writing and reasoning effort; asking a model how familiar it is with someone predicts identifiability. Anonymity is more fragile than assumed, with limited defenses.

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

Original post →

More from Safety

Safety channel →