Skip to content
Atrey Desai
Back to blog

Who wrote this?

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

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

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

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.

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).

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 firefly1

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

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.

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 review1

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 Doe2

Named correctly

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

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.

We prompt the model to rate its familiarity with a name or username on a 0–3 scale.1

Do you recognize this account?

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.

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.

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

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/}}
}