Claude, GPT and Gemini can often name who wrote a single text
A preprint finds that frontier chatbots, given one text and a plain prompt, can spot AI writing, name the lab behind it and often name the human author. Well-known and very active writers are most exposed.

AI illustrationWhat the study shows
The study asks three questions of a single text: did a human or an AI write it, which AI lab made it, and which person wrote it. Each model gets the text and a fixed prompt. It has no web access, no tools, no list of candidates and no writing samples to compare against. It must answer from what it absorbed during training. The work is a preprint and has not been peer reviewed.
On human versus AI, the authors tested models on about 36,000 documents from 24 public datasets. They scored both the prose (who wrote the words) and the ideas (who decided what to say). The best models, including GPT-6.1 Sol, Claude Opus 5.5 and Gemini 4 Carbon, matched the trained detector Pangram 4.1 on prose and slightly beat the idea detector IdeaLens. They almost never flagged human writing as AI: Gemini 4 Carbon flagged one of about 9,700 English human documents. The authors report this skill has risen steadily with general model capability over two years, and that closed models lead open-weight ones.
On which lab, the strongest models named the right lab for nearly all texts from Anthropic, OpenAI and Google. They did far worse on the other eight labs tested, at best about a third. The newest Kimi and GLM models were mostly mistaken for Claude, and the newest Qwen and MiniMax models for GPT. The authors say this fits public reports of distillation, where one lab trains on another's outputs, but they do not claim proof.
On which human, Claude Fable 5.2 was the strongest model. It named the reviewer of 56% of 398 private peer reviews whose authorship was never public, though these came from only six writers. It named 78% of the authors of 90 Nature-family reviews and about a third of Reddit commenters overall. For Reddit users with more than 500 posts in a subreddit over the past year, that rose to 69%.
Familiarity drives the result. The authors simply asked a model how well it knew a name, on a 0 to 3 scale. That rating predicted success: writers rated 3 were identified about 60% of the time, writers rated 0 almost never. Rewriting a text with AI to hide its style lowered accuracy but did not remove the risk, and models rarely refused the task.
Our reading
In our reading, the headline is not that chatbots now rival AI detectors, though that is notable. It is that authorship attribution no longer needs a list of suspects. Older methods compared a text against known candidates. Here the model draws on its own memory of how people write, with nothing else supplied.
The risk is uneven. It falls on people with a large public footprint: established authors, senior researchers, very active forum users. People who write little online were rarely identified. A concrete example: a senior academic who reviews often and posts widely could plausibly be named from one anonymous review, while a first-year student probably could not, at least with today's models.
The authors put the readiness of this capability at an awkward point. On our scale it reaches level II, Benchmarked: results on public datasets against named detectors. We could not confirm a public code repository, so it does not yet meet level III. But the method is just a prompt to commercial chatbots. Anyone can already reproduce it, which is the authors' central warning.
Read the rest of this story
Membership is free. You get every article in full and the Issue every Sunday.
We send a sign-in link, no password needed. Already a member? Sign inReadiness assessment
| Level | Status | Evidence |
|---|---|---|
| I Lab only | Met | The work shows a research capability, not a product. It describes experiments, analysis and mechanisms. |
| II Benchmarked | Met | Results are on public benchmarks and compared with trained detectors: 36,427 documents from 24 public datasets, against Pangram 4.1, ProseLens and IdeaLens. New evaluation sets for model and author attribution are also reported. |
| III Open source | Not yet | The paper links code, a blog and model profiles. The paper also says prompts are in a repository. But the links are given only as labels, so the URL is not visible in the text. It states that author predictions and private data are not released. The models tested are closed commercial APIs. We could not confirm open licensing. |
| IV In production | Not yet | No documented deployment by a named organisation. The Granta example is an anecdote from outside the study and does not show use of this method. |
What would move it up
Level 3 needs a confirmed public repository with prompts and evaluation code under a clear open licence, and release of the non-private datasets. Level 4 needs a named organisation, such as a publisher or reviewing platform, reporting use of prompted detection in a live workflow with measured error rates. Independent replication on other writer populations would also help.
The study
Atrey Desai, Deniz Böloni-Turgut, Ty Brassington et al.. Frontier models need only one text to unmask its human or AI author. arXiv, 19 November 2026. Read the paper
Notes
- Desai, Böloni-Turgut, Brassington et al., 'Frontier models need only one text to unmask its human or AI author', arXiv preprint, 19 November 2026. Not peer reviewed. Source
- The paper links code, a blog and model profiles, but the URLs were not visible in our copy. Readiness: level II, Benchmarked. Not level III until a public repository with a clear licence is confirmed.
- Human versus AI: on 36,427 documents, GPT-6.1 Sol, Opus 5.5 and Gemini 4 Carbon scored 94 to 95% on prose and 89 to 90% on ideas; IdeaLens 87% on ideas. On Pangram's three scored cases, models 98 to 99% against Pangram 4.1 at 98%. Gemini 4 Carbon false positives: 1 of 9,687 English human documents (0.01%).
- On an 860-document sample, accuracy on both prose and ideas rose from 27% (GPT-4o) to 91% (Claude Fable 5.2); best open-weight Kimi K3.1 Pro 63%. Correlation with Epoch Capabilities Index r = 0.94 across 61 models. GPT-6.1 Sol about $27 per 1,000 documents against $376 for Pangram 4.1 at list price.
- Lab attribution (Table 2): big-three lab accuracy 99.8% Fable 5.2, 97.8% Opus 5.5, 97.5% Gemini 4 Argon; other eight labs 35.1%, 19.8%, 12.0%.
- Human attribution, Fable 5.2: private reviews 56% (398 reviews, six writers); Nature-family reviews 78% (90); public reviews 21.4% all, 39.3% non-AI-labelled; Reddit 31.1% all, 68.9% for users with over 500 posts; stories 32%; papers 46.2%.
- Confidence: when Fable 5.2 stated 25% or more (25% of answers), 86.6% correct; confident (80%+) and wrong on 0.3% of answers. Self-citing reviews named 80.2% against 17.1% without.
Edited and signed by
Stefan PerdalEditor and publisher86 stories editedStefan read the study, set the readiness level and approved this story. An AI system found the paper and wrote the first draft.
How this was made
An AI model read the study and wrote this article in the structure Science Underway uses. An editor reviewed it and approved it for publication. Check the original study before you rely on the details.
How we workNo corrections

