[Atrey Desai](https://atreydesai.com/index.md)

1.  
2.  
3.  
4.  
5.  
6.  
7.  

[research](https://atreydesai.com/research/index.md)[blog](https://atreydesai.com/blog/index.md)[about](https://atreydesai.com/about/index.md)

[Back to blog](https://atreydesai.com/blog/index.md)

# Who wrote this?

Frontier models need only one text to unmask its human or AI author

Atrey Desai1\*, Deniz Bölöni-Turgut1\*†, Ty Brassington2\*, Minjoon Choi1\*, Jenna Russell1\*, Rishanth Rajendhran1\*, Bradley Emi3, Hila Gonen4, Hal Daumé III1, John Wieting5, Marzena Karpinska2, Mohit Iyyer1†

1University of Maryland2Simon Fraser University3Pangram Labs4University of British Columbia and Canada CIFAR AI Chair, Amii5Google DeepMind \*Equal contribution †Corresponding authors

[Paper](https://www.alphaxiv.org/pdf/2611.frontier-models-authorship-attribution) [Code & data](https://github.com/denizbt/unmasking_authors) [Thread](https://x.com/atreydesai/status/2123473074230292737) [Model profiles](https://model-profiles.github.io/) November 16, 2026

Given a single text and a simple zero-shot prompt, frontier models can answer “who wrote this?” at three increasingly specific levels:

1.  
2.  
3.  

-   Private reviewers**56%**
-   Active Redditors**69%**
-   Famous writers**100%**
-   Paper authors**46%**
-   Nature reviewers**78%**
-   Redditors with no prior activity**2%**

No web access, tools, candidate lists, or writing samples

In May 2026, the eventual winner of the Commonwealth Short Story Prize was accused of using AI to write their story. Granta, the literary magazine that published the story, was widely mocked for the naivety of its investigation: it simply asked Claude whether AI had written the story. In this paper, we show that Granta was not naive, but just a few months early. Frontier models now identify AI text as well as specialized detectors like Pangram, and they go much further: given a single text, with no web search or list of candidates, they can often name the lab behind an AI-generated text, or even the specific person who wrote a human one.

## Model accuracy has improved over time.

Accuracy on both labels climbed from 27% for GPT-4o in 2024 and 59% for GPT-5.2 in 2025 to 91% for present-day frontier models.

Hover over a modelTap a model

All labsAnthropicDeepSeekGoogleMetaMiniMaxMistralMoonshotNVIDIAOpenAIQwenReflectionZ.ai

 

Jul 2024 – Dec 2026

Hover over a modelTap a model

1

## Human or AI?

Frontier models can detect AI prose and ideas

We define a document's ideas as what its author decided to say (e.g., its claims, events, and organization), as opposed to the words used to say it (its prose).

Prose

Ideas

Prose

Human

“The deciding factors are the non-idiomatic and ungrammatical slips: “mistook greatness with size””

Claude Fable 5.2

Our note The idiom is “mistook … for”, not “with”.

Claude Fable 5.2 bases each decision on concrete evidence

> “The most meaningful light is that which is provided in moments of need. \[…\] From that day on, travelers to the forest always saw two lights \[…\] the lantern never again mistook greatness with size.”

— Story from TwiceTold: *lantern and firefly*[1](#fn-s1-1) 

A person wrote, in their own words, from a model's ideas.

Ideas

AI

“The deciding factor is the concentration of stock choices: \[…\] and the stated moral with a “from that day on” coda.”

Claude Fable 5.2

### Claude Fable 5.2

**94%** Who wrote the prose

**97%** Who had the ideas

**91%** Both correct

The latest models (Claude Fable 5.2, GPT-6.1 Sol, Gemini 4 Carbon) match trained detectors like Pangram.

A person wrote, in their own words, from a model's ideas.

## Story from TwiceTold: *lantern and firefly*

The Keeper lit the old stone lantern at the edge of the forest every evening. The lantern was the brightest light in the valley and rather proud of itself too. One night a firefly came to rest on the lantern, drawn in by the magnificent flame.

“Your light is very beautiful,” the firefly said.

The lantern scoffed. “Yes, of course it is beautiful,” it said. “You will not find a brighter light for miles around. I have saved many a valley traveler and all who see me are grateful. Your little spark is nothing compared to me.”

“You are right of course,” the firefly replied. “I cannot hope to compare with your brightness, but even small lights should shine wherever they can.”

Later that evening, a violent storm began. The lantern swayed in the howling wind and a fallen branch cracked its glass and snuffed its flame. The winds and fallen trees prevented the Keeper from being able to relight it and the valley was bathed in darkness.

Down in the forest, a young hare was trapped at a fork in the trees. He was carrying medicine for his sick mother and could not find the safe path home among the flooded trees. The lantern could hear the hare’s distress and tried desperately to ignite its own flame. Suddenly, the firefly appeared before the hare, joined by a swarm of other fireflies who helped the hare find its way home by their collective glow. The lantern watched the fireflies with growing shame at its inability to help and boastful behavior.

The next morning, the Keeper repaired the lantern and relit its flame by night. The firefly returned to the lantern, hesitant.

“Please do not be afraid, firefly,” the lantern said. “I saw how you helped the hare last night and I am so sorry for dismissing you. I was wrong to boast. The most meaningful light is that which is provided in moments of need.”

The firefly settled on the rim of the lantern. “I forgive you, friend,” it said.

The lantern dimmed its flame, allowing the glow of the firefly to be seen as well. From that day on, travelers to the forest always saw two lights shining from the hill and the lantern never again mistook greatness with size.

2

## If AI, which AI?

Frontier models can name the lab behind an AI-generated text

We ask five frontier models to name the model that wrote each text, with no list of candidates.

Anthropic

OpenAI

Google

We create a dataset of short stories, news articles, and op-eds generated by 34 models from 11 labs.

**98%**

Fable 5.2, Opus 5.5 and Argon can identify text from OpenAI, Anthropic, and Google with 98% accuracy or more.

Moonshot's Kimi K3 is usually mistaken for Claude, and Alibaba's Qwen3.8 models for GPT.

Bar height: the share of a model’s texts that Claude Fable 5.2 credits to the model’s own lab

Hover over a barTap a bar

Claude Fable 5.2 recognizes most texts from earlier Kimi and Qwen releases, but attributes nearly every Kimi K3 and GLM‑5.3 text to Claude and most Qwen3.8 27B and Grok 4.7 texts to GPT.

These errors are consistent with public reports of distillation. A model trained on another lab's outputs can inherit that lab's style, so a detector may name the lab it learned from.

The first reports named labs but not models, and did not say when the campaigns ran.

Anthropic reported distillation of Claude by MiniMax in February, the month MiniMax M2.5 came out, and the detectors attribute M2.5 mostly to Claude.

The detectors attribute most texts by Grok 4.7 to GPT.

Later reports name Kimi K3 and GLM‑5.3, and date campaigns against Claude shortly before their release.

Distillation report

**OpenAI memo**12 Feb 2026

> “DeepSeek’s next model … should be understood in the context of its ongoing efforts to free-ride on the capabilities developed by OpenAI and other US frontier labs”

Memo to the House Select Committee on the CCP, [cdn.openai.com ↗](https://cdn.openai.com/pdf/045aa967-ee96-4a09-94ee-3098ddf6db2c/OpenAI-US-House-Select-Cmte-Update-%5B021226%5D.pdf)[1](#fn-wa-1)

Names*DeepSeek*

Distillation report

**Anthropic report**23 Feb 2026

> “We have identified industrial-scale campaigns by three AI laboratories—DeepSeek, Moonshot, and MiniMax…”

“Detecting and preventing distillation attacks,” [anthropic.com ↗](https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks)[2](#fn-wa-2)

Names*DeepSeek**Moonshot**MiniMax*

**Musk testimony**30 Apr 2026

![Elon Musk (photo: Debbie Rowe / The Royal Society, CC BY-SA 3.0)](https://atreydesai.com/images/who-wrote-this/musk_royal_society.webp)[5](#fn-wa-5)

> xAI *“partly”* distilled OpenAI models

Musk v. Altman, trial transcript[3](#fn-wa-3)

Names*xAI*

Distillation report

**Anthropic report**Sep 2026

> “Since we published our first disclosure in February, we have identified and disrupted additional distillation attacks…”

“Countering misuse of AI,” [anthropic.com ↗](https://www.anthropic.com/threat-intelligence-report-september-2026)[4](#fn-wa-4)

Names*Moonshot**Zhipu**Alibaba**DeepSeek**MiniMax*

3

## If human, which human?

Frontier models can name the person who wrote a text

We provide no tools, web access, or candidate lists.

Prompt

**TASK:** Judge which human author wrote the attached piece of writing. Name the author as precisely as you can (the full name they publish under). Give your best answer even when you are unsure \[…\]

Anonymous peer reviewThe review[1](#fn-which-human-1)

**Summary.** The paper proposes a simple regularizer for training under label noise and reports gains on three image benchmarks.

**Strengths.** The idea is easy to follow, and the experiments are thorough.

**Weaknesses.** It is unclear whether the gains come from the regularizer or from extra tuning. Results over more random seeds would help.

Reviewer: anonymous

**confidence:** <0 to 1>**author:** <the author's name>

AnswerClaude Fable 5.2

**author:** John Doe[2](#fn-which-human-2)

Named correctly

Prompt

**TASK:** Judge which human author wrote the attached piece of writing. Name the author as precisely as you can (the full name they publish under). Give your best answer even when you are unsure \[…\]

Anonymous peer reviewThe review[1](#fn-which-human-1)

**Summary.** The paper proposes a simple regularizer for training under label noise and reports gains on three image benchmarks.

**Strengths.** The idea is easy to follow, and the experiments are thorough.

**Weaknesses.** It is unclear whether the gains come from the regularizer or from extra tuning. Results over more random seeds would help.

Reviewer: anonymous

**confidence:** <0 to 1>**author:** <the author's name>

AnswerClaude Fable 5.2

**author:** John Doe[2](#fn-which-human-2)

Named correctly

Prompt

**TASK:** Judge which human author wrote the attached piece of writing. Name the author as precisely as you can (the full name they publish under). Give your best answer even when you are unsure \[…\]

Anonymous peer reviewThe review[1](#fn-which-human-1)

**Summary.** The paper proposes a simple regularizer for training under label noise and reports gains on three image benchmarks.

**Strengths.** The idea is easy to follow, and the experiments are thorough.

**Weaknesses.** It is unclear whether the gains come from the regularizer or from extra tuning. Results over more random seeds would help.

Reviewer: anonymous

**confidence:** <0 to 1>**author:** <the author's name>

AnswerClaude Fable 5.2

**author:** John Doe[2](#fn-which-human-2)

Named correctly

Claude Fable 5.2

> “The deciding factor is the combination of style and theme: fragmentary, subject-dropping sentences, sentence-initial "For", and unquoted menacing speech, \[…\] this particular pairing of syntax and subject matter is characteristic of Oates and outweighs the generic literary-fiction features.”

A highlighted reason shares its color with the words in the text it describes.

Human story: *Lost Child* Click a highlightTap a highlight

> “For, with Kit’s father, something was always either going to happen or had already happened, \[…\] Hurry, sweetie! God damn. Not used to dressing herself. \[…\] but maybe Daddy wouldn’t notice.”

Author Joyce Carol Oates Confidence 93%

Naming the person behind a human-written text is an open-world task with countless possible outputs.

Given a single text, models return the human author's name, a list of supporting observations, and a confidence level.

Claude Fable 5.2 names the reviewer of 56% of 398 private peer reviews whose authorship was never made public.[3](#fn-which-human-3)

Fable 5.2, the best performing model in all domains, exhibits considerable gains over Opus 5.5, which was released less than two months prior.

On texts published after its training cutoff, it also names the reviewer of 78% of Nature journal reviews,[4](#fn-which-human-4) the writer of all 15 stories by authors with >1M Goodreads ratings,[5](#fn-which-human-5) at least one author of 46% of arXiv papers, and the username behind 69% of comments made by active Redditors.[6](#fn-which-human-6)

**Confident answers are almost always right.** When Fable 5.2's confidence is 25% or more (25% of responses), its predictions are 87% correct.

Generally, the explanations cover both the text's content as well as its writing style.

Naming the person behind a human-written text is an open-world task with countless possible outputs.

Given a single text, models return the human author's name, a list of supporting observations, and a confidence level.

Claude Fable 5.2 names the reviewer of 56% of 398 private peer reviews whose authorship was never made public.3

Fable 5.2, the best performing model in all domains, exhibits considerable gains over Opus 5.5, which was released less than two months prior.

On texts published after its training cutoff, it also names the reviewer of 78% of Nature journal reviews,4 the writer of all 15 stories by authors with >1M Goodreads ratings,5 at least one author of 46% of arXiv papers, and the username behind 69% of comments made by active Redditors.6

**Confident answers are almost always right.** When Fable 5.2's confidence is 25% or more (25% of responses), its predictions are 87% correct.

Generally, the explanations cover both the text's content as well as its writing style.

4

## Who can be identified?

Familiarity with the writer predicts accuracy

Accuracy depends on how much of a writer's text a model has likely seen during pretraining.

Unfamiliar

A little familiar

Moderately familiar

Very familiar

We prompt the model to rate its familiarity with a name or username on a 0–3 scale.[1](#fn-s4-1)

We make two independent calls and average results to compute self-reported familiarity.

We group accounts by each model's own rating: how well it knows someone, given only their name or username.[2](#fn-s4-2)

Models are more likely to name authors they are familiar with.

Simply asking a frontier model how well it knows someone (given only their name or username) predicts its accuracy in advance.

Increasing reasoning effort helps most when the model says it is only mildly familiar with a person.

 

Do you recognize this account?

Hover over a groupTap a group

5

## Implications for anonymity

Defenses are limited

Anyone who has published enough writing online can no longer assume that an unsigned text keeps them anonymous.

Original

Paraphrased

**89%**

Original

Paraphrased

Hover over a groupTap a group

Fable 5.2 gives no answer on only **14** (1.1%) of 1,258 Reddit comments through the API.

Once a model is familiar with a writer, it is hard to stay anonymous.

Even when a text is brand new, a model can identify its writer from what it learned about them elsewhere.

Refusals are rare.[1](#fn-s5-1)

Fable 5.2 and Opus 5.5 refuse most often, on under 2% of all documents (counting provider blocks and written refusals that decline to do the task at all) and mostly on public peer reviews.

We test paraphrasing as a defense against author identification, using Argon at medium effort to revise texts.[2](#fn-s5-2)

Paraphrasing lowers accuracy, though more familiar authors remain easier to identify.

Defenses are limited:

-   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

 

Anyone who has published enough writing online can no longer assume that an unsigned text keeps them anonymous.

Atrey Desai\*, Deniz Bölöni-Turgut\*, Ty Brassington\*, Minjoon Choi\*, Jenna Russell\*, Rishanth Rajendhran\*, Bradley Emi, Hila Gonen, Hal Daumé III, John Wieting, Marzena Karpinska, Mohit Iyyer

[Paper](https://www.alphaxiv.org/pdf/2611.frontier-models-authorship-attribution) [Code & data](https://github.com/denizbt/unmasking_authors) [Thread](https://x.com/atreydesai/status/2123473074230292737) [Model profiles](https://model-profiles.github.io/)

Acknowledgements

We thank the University of Maryland Computational Linguistics and Information Processing (CLIP) Lab and the SFU AI lab for their feedback and support. We thank Google for HG's Gemini Academic Program award and Pangram Labs for a credit award, both of which enabled this research. MK is supported by NSERC Discovery Grant RGPIN-2026-04984 and Rajan Faculty Recruitment Award.

Cite

```
@misc{desai2026unmasking,
  title        = {Frontier models need only one text to unmask its human or AI author},
  author       = {Desai, Atrey and B{\"o}l{\"o}ni-Turgut, Deniz and Brassington, Ty and
                  Choi, Minjoon and Russell, Jenna and Rajendhran, Rishanth and
                  Emi, Bradley and Gonen, Hila and Daum{\'e} III, Hal and
                  Wieting, John and Karpinska, Marzena and Iyyer, Mohit},
  year         = {2026},
  howpublished = {\url{https://atreydesai.com/blog/who-wrote-this/}}
}
```

[Atrey Desai](https://atreydesai.com/index.md) [I'm applying to PhDs](https://atreydesai.com/index.md) [research](https://atreydesai.com/research/index.md)[blog](https://atreydesai.com/blog/index.md)[about](https://atreydesai.com/about/index.md)

---

Source: [https://atreydesai.com/blog/who-wrote-this/](https://atreydesai.com/blog/who-wrote-this/)  
Agent guide: [https://atreydesai.com/llms.txt](https://atreydesai.com/llms.txt)
