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

Atrey Desai, Deniz Böloni-Turgut, Ty Brassington, Minjoon Choi, Jenna Russell, Rishanth Rajendhran, Bradley Emi, Hila Gonen, Hal Daumé, John Wieting, Marzena Karpinska, Mohit IyyerarXiv 2026 · arXiv

TL;DR

Frontier large language models (LLMs) like Claude Fable 5.2 and GPT-6.1 can now identify the author of a single text snippet—whether AI or human—using only their internal parametric knowledge. These models match specialized AI detectors and can deanonymize human authors (including "anonymous" peer reviewers and Redditors) with startling accuracy, posing a significant new threat to online privacy.

What Problem Does This Paper Address?

Authorship attribution—the task of determining who wrote a specific text—has traditionally required specialized tools, large reference corpora, or a defined list of candidates. Similarly, detecting AI-generated text has been a "cat-and-mouse" game involving trained classifiers that often struggle with new model versions.

This paper investigates whether "frontier" models (the most capable LLMs available) have reached a tipping point where they can perform these tasks zero-shot. That is, given a single text and no external aids (no web search, no candidate lists), can they identify if a human or AI wrote it, which specific AI lab produced it, or even the specific identity of a human author? The authors explore the implications of these models' vast "parametric memory"—the stylistic patterns they've absorbed during pretraining—for the future of anonymity.

Methods

The researchers evaluated 72 models from 13 labs released between 2024 and 2026. They employed a simple, zero-shot prompting strategy across three increasingly granular tasks:

  1. Human or AI? Distinguishing between human prose/ideas and AI-generated content.
  2. If AI, which lab? Identifying which of 11 different labs (e.g., OpenAI, Anthropic, Google) produced the text.
  3. If human, which human? Naming the specific individual or username responsible for the text.

The study also introduced a "familiarity" metric, where models were asked to rate their own knowledge of a specific author on a 0–3 scale, and analyzed how "reasoning effort" (e.g., chain-of-thought) affected attribution accuracy.

Figure 1. Given a text sample with no candidates or external aids, Claude Fable 5.2 can readily identify AI authorship, source labs, and even human writers. Frontier models still face challenges in identifying open-weight models and less-known authors. Open models: percent of texts by open-weight writers that Fable 5.2 attributes to any open-weight model; naming the exact lab is harder (Table 2).

Datasets

The evaluation utilized a massive collection of over 37,000 documents across 24 public benchmarks and several newly curated datasets:

  • AI Detection: Benchmarks like MELD and StoryScope were used to test prose and idea detection.
  • Model Attribution: A new dataset of stories, news, and op-eds generated by 34 models from 11 labs.
  • Human Deanonymization:
    • Private Data: 398 peer reviews written by the paper's co-authors (never published publicly).
  • Recent Public Data: 1,559 peer reviews from 2026 (post-training cutoff), 120 recent arXiv papers, and 1,258 long-form Reddit comments from 32 subreddits.

Key Insights and Ideas

  • Parametric Fingerprinting: Frontier models don't need to "search" the web; they have essentially memorized the "stylistic fingerprints" of both AI labs and prolific human writers.
  • The "Big Three" Dominance: Models are nearly perfect (98%+) at identifying text from OpenAI, Anthropic, and Google. However, they often misattribute text from smaller labs (like Moonshot or Alibaba) to the "Big Three," likely due to model distillation, where smaller models are trained on the outputs of larger ones and inherit their style.
  • Familiarity is the Best Predictor: A model's self-reported familiarity with a name is a highly reliable indicator of whether it can identify that person's writing. If a model says it knows you "well," your anonymity is likely compromised.
  • Anonymity is Fragile: Even if a specific text (like a private peer review) was never in the training set, the model can identify the author based on their other public writings. This makes double-blind review increasingly difficult to maintain.
Figure 6. Accuracy increases significantly for more active Redditors and authors with higher visibility (Goodreads ratings).

Key Results

  • AI Detection: Gemini 4 Carbon matched the state-of-the-art specialized detector (Pangram 4.1) with 97% accuracy on fully AI-generated texts.
  • Lab Attribution: Claude Opus 5.5 correctly identified the source lab for 98% of texts from the major providers.
  • Human Deanonymization:
    • Claude Fable 5.2 identified the author of 78% of public reviews from the Nature family.
  • It correctly named the username behind 69% of comments from active Redditors (>500 posts/year).
  • Even on private, never-before-seen peer reviews, it reached 56% accuracy.
  • Confidence Calibration: When Fable 5.2 reported a confidence level of 25% or higher, its predictions were correct 87% of the time.
Figure 2. Accuracy of prose and idea detection has steadily improved with model release dates, with closed-weight models consistently leading.

Limitations and Future Work

  • Open-Weight Models: Models still struggle to identify the specific lab for open-weight models (only ~55% accuracy), as these models often lack the distinct "corporate" style of the major labs.
  • The "Long Tail" of Authors: While famous or highly active writers are easily unmasked, the models struggle with authors who have a small online footprint.
  • Defenses are Weak: The authors tested paraphrasing as a defense; while it lowered accuracy, it did not eliminate the risk, especially for authors the model "knew" well.
  • Future Directions: The researchers suggest that model developers should implement "refusal" mechanisms for authorship queries and that the scientific community needs to rethink the feasibility of double-blind peer review in an age of frontier LLMs.

Summary updated

Abstract

Given a single text and a simple zero-shot prompt (no web access, tools, candidate lists, or writing samples), frontier models can answer "who wrote this?" at three increasingly specific levels: (1) whether an AI or human wrote it, (2) if an AI, which lab, and (3) if a human, often exactly who. The latest models (Claude Fable 5.2, GPT-6.1 Sol, Gemini 4 Carbon) match trained detectors like Pangram while also explaining their predictions in natural language, and Opus 5.5 names the correct lab for 98% of texts generated by models from Anthropic, OpenAI, and Google. Most strikingly, Fable 5.2 correctly names the author of 39% of public peer reviews published in Oct-Nov 2026 (78% on a subset of reviews from Nature), and identifies the username behind 69% of comments from active Redditors (with >500 posts over the last year). Accuracy depends on exposure to an author's writing in pretraining data and on reasoning effort, and simply asking a model how familiar it is with someone predicts whether it can identify their writing. For anyone who has published enough writing online, frontier models present a greater threat to anonymity than commonly assumed, and defenses are limited.
BibTeX
@misc{desai2026frontier,
  author = {Atrey Desai and Deniz Böloni-Turgut and Ty Brassington and Minjoon Choi and Jenna Russell and Rishanth Rajendhran and Bradley Emi and Hila Gonen and Hal Daumé and John Wieting and Marzena Karpinska and Mohit Iyyer},
  title = {{Frontier models need only one text to unmask its human or AI author}},
  year = {2026},
  eprint = {2611.26216},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.26216},
}
Figure