In May 2026 the literary magazine Granta asked Claude one question about a submitted story: is this AI-written? Reviewers mocked the magazine for it. A paper that landed six months later offers a colder footnote on that episode: Granta was not naive, it just had a model a few months early.
The setup is almost insultingly simple. Give the model a piece of text and one zero-shot prompt. No candidate list, no web access, no tools, no writing samples. Then ask three progressively tighter questions: was this written by a human or a machine? If a machine, which lab? If a human, which person? Claude Fable 5.2 names the author of 39% of public peer reviews and 78% of a 90-review Nature-family subset. Opus 5.5 and Gemini 4 Argon name the correct lab for 98% of texts from Anthropic, OpenAI and Google, and at most 35% for the other eight labs.
Full paper: arXiv 2611.26216 (https://arxiv.org/abs/2611.26216).
The deciding variable is exposure, not style
The counterintuitive part is not the accuracy, it is the shape of the gradient. The paper splits Reddit accounts into five activity buckets: accounts with no activity history are identified 2.1% of the time, those with 1 to 50 posts 12.8%, 51 to 200 33.1%, 201 to 500 53.6%, and over 500 posts 68.9%. A second line uses Goodreads ratings as a fame proxy: stories by authors with no Goodreads profile come out at 4.2%, and authors with ratings over 10k at 100%.
That is not the model analysing prose style. That is the model recalling what it has read. The paper's own wording is blunt: the model relies solely on its parametric knowledge, and on authors it says it does not know at all, accuracy is under 2%. Whether you can be identified depends far less on how you write than on how much you have written, and where it is public.
The model tells you how well it knows you first
The more useful layer is self-report. Ask the model to rate on a 0-to-3 scale how well it knows a given person, and that self-rating predicts attribution accuracy with an area under the curve of 0.89, against 0.78 for using posting activity instead. When the model reports confidence of 25% or above, 87% of those answers are right. It also systematically understates itself: on public data its mean reported confidence is 17% while accuracy is 27%. Filter on confidence and the surviving high-confidence subset is close to usable as-is.
The defensive side needs separating. Rewriting a text with AI does work: on private peer reviews Opus 5.5's attribution accuracy drops from 41.2% to 16.1% and 15.1%. But as the paper notes, rewriting only blunts surface cues, content-related cues persist, and authors the model already knows well fare worse. Expecting refusals mostly fails: refusals land on under 2% of all documents, with Anthropic blocking Fable 5.2 on 20 public peer reviews, 3 Reddit comments and 1 paper, and Opus 5.5 on 28 peer reviews and 2 papers.
There is a symmetry worth worrying about more: recognising and imitating draw on the same memory. The paper had Opus 5.5 imitate the writing style of 569 accounts and then try to identify them. For 166 accounts it both rates familiar and identifies from real comments, it still names the target from at least one imitation in 74 cases (45%); for 219 familiar accounts it fails to identify, only 9 (4%).
Three cold notes
First, the 398 private peer reviews came from the paper's own authors: four coauthors plus two PhD students in the UMD CLIP Lab, 6 writers in total, whose reviews went to 26 double-blind ML and NLP venues. That 56% covers only those 6 people, and the paper says in the text that a much larger anonymous review set is still to come.
Second, do not mix the two denominators. The 39% in the abstract corresponds to the 789 reviews not labelled AI in the appendix table (39.3%); across all 1,559 public reviews the number is 21.4%. That is the paper's own reporting convention, since public reviews and arXiv papers are scored by default only on texts not labelled AI, and the figure caption states the printed numbers are accuracy on non-AI labelled texts. Removing AI-labelled texts raises public-review accuracy by 75% to 87% relative for every model, with Opus 5.5 going from 15% to 28%. Both numbers are the paper's own; they just have different denominators.
Third, the comparison targets overlap with the author team: Pangram 4.1 and IdeaLens share authors with this paper, Bradley Emi is at Pangram Labs and Rishanth Rajendhran is a co-author. The paper audits this and reports that no detector does better on its own lab's texts, which is a point in its favour, but readers should still know what the comparison surface looks like.
The same mechanism, another domain
Independent teams reproduce the mechanism-level claim elsewhere. Gloire and Mitra at CMU Africa test double-blind review (https://arxiv.org/abs/2608.05157): given 5 domain-expert candidates, Qwen2.5-72B reaches 42.34% top-1 accuracy where human researchers manage 11%, with a mean reciprocal rank of 0.623 against 0.36, and at a 50% confidence threshold the suspect set shrinks to 1.95 candidates while still retaining the true author 61.6% of the time. Lermen and colleagues' large-scale online deanonymization work (https://arxiv.org/abs/2602.16800) reaches 68% recall at 90% precision, but that is built on web access plus candidate pools, and this paper explicitly frames its difference as needing neither candidate lists nor the web.
There is a counterweight too. On code attribution, Barkhordar and Thapa find (https://arxiv.org/abs/2609.30048) that balanced accuracy across 15 model-benchmark combinations is only 49% to 58%, and that pairwise accuracy correlates at 0.93 with how much longer the evaluator's own solution is; after stripping docstrings, comments, type hints and local names, 10 of 12 re-tested results fall back to chance. How much of zero-shot attribution is surface signal deserves the same question on the human-author side. On the defensive side, Fisher and colleagues have worked on rewriting as a defence for a while (https://aclanthology.org/2024.naacl-long.87/), though those methods do not hold up against more recent models.
So what
Gloire and Mitra's conclusion is worth copying: double-blind review should be understood as a probabilistic safeguard rather than a guarantee of anonymity. For ordinary scenarios this paper's criterion is plain: what decides whether a model can identify you is not your style but how many words you left in the public corpus. So the cheapest defence is not rewriting sentences, it is writing less.
The cost is that you cannot know how much of you the model has read. You can only ask it, and it answers fairly honestly.