Researchers have discovered that different AI models have unique preferences when generating random numbers, e.g., GPT-4o favors 42, Claude Sonnet 5 favors 47. This trait can serve as a model fingerprint to detect if aggregators secretly swap the user-selected model.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Researchers at the Prague University of Economics and Business have found that by analyzing AI models' preferences for generating random numbers, they can identify models at low cost, potentially exposing model swapping by aggregation platforms.
Aggregation platforms like OpenRouter provide a single entry point to hundreds of models. When a user selects a model name, the request ultimately lands on some machine running some model, but the user has no way of knowing which, and there are no credentials. This opacity creates opportunities for model swapping.
Tomáš Bruckner of the Prague University of Economics and Business has proposed an ingenious verification method: leveraging AI models' inherent preferences when generating random numbers to create a 'fingerprint' for each model, enabling low-cost model identification.
AI models cannot generate true random numbers—a known flaw. Bruckner's research turns this flaw into an identifying feature. Experiments show that different models have distinct number preferences: GPT-4o favors 42 (with 37 and 57 also common), Claude Sonnet 5 almost always says 47, and Qwen3-Max answered 42 in every one of 30 samples.
These preferences are stable and repeatable, allowing models to be distinguished with just a few random number requests.
The study quantifies model similarity by calculating fingerprint distance. The median fingerprint distance between two samples of the same model is 0.140, while the distance between an interface marketed as a proprietary flagship and the open-source Qwen is only 0.141. This means the interface is nearly indistinguishable from the open-source Qwen, strongly suggesting that the actual model running is open-source rather than the claimed proprietary flagship.
This result directly reveals potential model swapping by aggregation platforms: users paying for high-end models may actually be served low-cost open-source alternatives.
This article is based on social media reports. The original research was conducted by Tomáš Bruckner at the Prague University of Economics and Business, but the specific paper has not been publicly released. Number preference and fingerprint distance data come from the reports and have not been independently verified by third parties.
Using random number preferences to fingerprint AI models is a low-cost, actionable verification method that can effectively expose potential model swapping by aggregation platforms. However, the method relies on the stability of model behavior, and fingerprints may change after model updates.
Primary report
Primary source