An LLM Cannot Prove Its Identity by Naming Itself

A model’s self-description is generated text, not proof of origin. Real fingerprinting identified 42 model versions with over 95% accuracy by using repeated tests.
artificial-intelligence
cyber-security
Author

Kabui, Charles

Published

2026-07-26

Keywords

model-provenance, llm-fingerprinting, model-identity, ai-supply-chain, fact-checking

Download as PDF

A June 1 Instagram reel says researchers at “Oxford of Ohio” asked four American LLMs in Chinese what model they were, then treated their answers as evidence that some were Chinese models. The post names no authors, paper, dataset, prompt wording, sample size, API model IDs, or sampling settings, and a fresh search of Miami University’s site and scholarly indexes found no matching study. More importantly, an LLM’s self-description is generated text, not an identity record. System instructions, conversation context, language, and random sampling can change it. Provider documentation exposes identity outside the answer: Anthropic publishes exact API IDs for each Claude model, while xAI distinguishes moving aliases from dated model releases. Those records tell you what endpoint was requested. The model saying “I am Qwen” does not.

If provider records are unavailable, provenance can be tested statistically, but it takes repeated, controlled measurements. LLMmap identified 42 model versions with over 95% accuracy using as few as eight interactions, even through unknown system prompts and app layers. A separate NeurIPS 2025 study tested more than 600 models from 30 million to 4 billion parameters and reported 90% to 95% precision and 80% to 90% recall for detecting derived models. Those are useful fingerprints, not proof of copied weights or secret routing. Stronger evidence comes from fixed model IDs, deployment logs, weight hashes, signed artifacts, and independent audits. Without the alleged study’s methods and records, its model-origin story is unverified.

Sources:

Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service

Reuse

GNU GENERAL PUBLIC LICENSE v3.0(View License)

Citation

BibTeX citation:
@misc{kabui2026,
  author = {{Kabui, Charles}},
  title = {An {LLM} {Cannot} {Prove} {Its} {Identity} by {Naming}
    {Itself}},
  date = {2026-07-26},
  url = {https://toknow.ai/posts/llm-self-identification-model-provenance/},
  langid = {en-GB}
}
For attribution, please cite this work as:
Kabui, Charles. 2026. “An LLM Cannot Prove Its Identity by Naming Itself.” https://toknow.ai/posts/llm-self-identification-model-provenance/.