Original text(via HN AI)
- Atrey Desai@atreydesaiFrontier models need only one text to unmask its author. Give a frontier model a single text, and it can often tell whether AI wrote it, which lab's AI, or even the specific person behind it. No web search, no list of candidates, with one text. 🧵00:00Mohit Iyyer and 8 others5:46 PM · Nov 19, 2026·3,166Views2

- Atrey Desai@atreydesai1hAnyone who has published enough writing online can no longer assume that an unsigned text keeps them anonymous. Interactive explainer: atreydesai.com/blog/who-wrote… Paper: alphaxiv.org/pdf/2611.front…Who wrote this? Frontier models need only one text to unmask its human or AI author | Atrey DesaiFrom atreydesai.com1Atrey Desai@atreydesai1hThe result that should worry anyone who reviews anonymously: Claude Fable 5.2 names the reviewer of 56% of 398 private peer reviews written by our lab, never made public. That's six reviewers. We're now running it on thousands of anonymous peer reviews. Results soon 👀1Atrey Desai@atreydesai1hLevel 1/3: Is it AI? When Granta simply asked Claude whether AI had written a prize-winning story, the internet laughed. Granta was just a few months early. Frontier models now spot AI text as well as specialized detectors like Pangram.1Atrey Desai@atreydesai1hAnd not just AI prose. AI ideas too. GPT-6.1 Sol reaches 90% accuracy on idea detection, beating IdeaLens (87%). Unlike most detectors, frontier models can also explain their predictions in natural language.1Atrey Desai@atreydesai1hLevel 2/3: Which AI wrote it? Fable 5.2, Opus 5.5 and Argon can identify text from OpenAI, Anthropic, and Google with 98% accuracy or more. The twist of distillation: Moonshot's Kimi K3 is usually mistaken for Claude, and Alibaba's Qwen3.8 models for GPT.1Atrey Desai@atreydesai1hHow do they tell? Ask them. The explanations reveal how frontier models view each lab's style: • GPT hedges its claims and leans on aphorisms. • Claude writes with dry wit and ends quietly. • Gemini stacks adjectives and invents named sources and quotes.1Atrey Desai@atreydesai1hLevel 3/3: Which human wrote it? On texts published after its training cutoff, Fable 5.2 names: • the reviewer of 78% of Nature journal reviews • the writer of all 15 stories by authors with >1M Goodreads ratings • at least one author of 46% of arXiv papers1Atrey Desai@atreydesai1hAnd when it is confident, it is almost always right. When Fable 5.2's confidence is 25% or more, its predictions are 87% correct. Who is most exposed? Simply asking a frontier model how well it knows someone predicts its accuracy in advance.1Atrey Desai@atreydesai1hCan you hide? Not really. • Models rarely refuse these queries. • Deleting one's online writing does not remove it from the weights of existing models. • Rewriting human-authored text with AI only helps partially.2Atrey Desai@atreydesai1hThis is joint work with my wonderful co-first authors @denizbtbt, @TyBrassington, @mjoonixix, @jennajrussell, and @rishanthrajendh, and with @bradley_emi, @hila_gonen, @haldaume3, @johnwieting2, @mar_kar_, and @MohitIyyer.2Atrey Desai@atreydesai1hWe thank @ClipUmd and the @SFU AI Lab for their feedback and support, and @Google and @pangram for the awards that enabled this research.
- Kevin Ren @ NeurIPS 2026@kevinren091hSuper interesting! You could imagine LLMs complicate author attribution since they can be used to deanonymize (as you do) and anonymize writing (if I pass my writing through an LLM then my writing style is homogenized). I wonder if this works when LLM use in writing is widespread1
