Machine Learning

One Token Is Enough: Fingerprinting and Verifying Large Language Models from Single-Token Output Distributions

Tomas Bruckner

Featured July 17, 2026

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

By asking LLMs simple questions like "name a random number," their unique, non-random answer patterns act like a fingerprint, letting users cheaply check if an online AI is truly the model it claims to be.

In depth

The paper introduces a novel method for black-box verification and attribution of Large Language Models (LLMs) by analyzing their single-token output distributions. It demonstrates that LLMs exhibit reliably idiosyncratic and non-uniform responses to trivial one-word prompts, such as "name a random number." This behavioral fingerprint, collected at minimal cost, allows for robust identification of LLMs, recovery of their lineage, and detection of model substitutions in opaque API serving chains.

Key Takeaways

1
LLMs possess a unique behavioral fingerprint derived from their non-uniform single-token responses to simple prompts, enabling identification without access to internal model states.
2
A cost-effective verification protocol is established, achieving a 7.3% equal error rate with a full battery of 40 probe cells, requiring only hundreds of single-token queries per audit.
3
The method successfully recovers model lineage with 59.5% accuracy and reveals ecosystem anomalies, such as proprietary endpoints distributionally indistinguishable from open-weight models.

Conceptual Flow

HIGH LEVEL

Methodology: Using AI's Quirks to Identify It

The study asks different AIs simple questions and looks at their unique, quirky answer patterns to tell them apart, like recognizing friends by their habits.

Simple Questions

Many AIs

Collect Answers

Unique Answer Patterns

Results: Confirming AI Identity and Family

These unique answer patterns reliably show which AI is which, help group them into families, and expose when an AI pretends to be another.

Unique Answer Patterns

Compare & Group

AI Identity Confirmed

AI Family Tree

Fake AI Detected

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.