EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-ND 4.0

Using Behavioural AI to Fingerprint Large Language Models

Paul Böhm, Marek Opuszko, Simon Emde

International Conference on AI Research · 18.11.2026 · 7 (1) · 34-44

DOI: 10.34190/icair.7.1.5580 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Conference on AI Research, online erschienen am 18.11.2026. Verfasst von Paul Böhm und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Identifying which large language model (LLM) produced an answer usually relies on semantic prompts, network timing or special tokens. This paper studies a much simpler probe: asking a model for random numbers. We asked ten LLMs of three providers (OpenAI, Google and DeepSeek) to generate a sequence of 50 random numbers between 1 and 10, collected 800 valid answers per model and compared them with a pseudo-random number generator (PRNG). Over all positions, the numbers are almost perfectly uniformly distributed, so a simple frequency test cannot separate the models from a PRNG. The order of the numbers, however, is far from independent. Nine of the ten models avoid immediate repetition and small steps, and in more than 97% of their sequences the first ten numbers are all different, which resembles drawing from an urn without replacement. All ten models most often start with the number 7, and the entropy of the numbers drops at the end of every block of ten positions. These patterns differ between models and can serve as behavioral fingerprints. We trained a long short-term memory (LSTM) network, a random forest and a logistic regression to recognize the source among eleven classes (ten LLMs and the PRNG). To avoid leakage, identical sequences were removed before splitting, and results were cross-validated and reported with confidence intervals. One sequence of 50 numbers identifies the source with 81.4% accuracy (95% confidence interval 80.5% to 82.2%). Five independent requests of 25 numbers each reach 97.2% (96.3% to 97.9%), and five requests of only 15 numbers still reach 95.5% (94.2% to 96.5%). The most frequently confused sources are the two DeepSeek models (deepseek-v4.1-flash and deepseek-v4-pro) and, separately, gpt-3.5-turbo-0125 and the PRNG. The observations are black-box: they hold for one fixed prompt and a closed set of models and describe behavior, not its cause. The results also caution against using LLMs as sources of randomness.

Zitieren

APA

Böhm, P., Opuszko, M., & Emde, S. (2026). Using Behavioural AI to Fingerprint Large Language Models. International Conference on AI Research, 7(1), 34-44. https://doi.org/10.34190/icair.7.1.5580

Vancouver

Böhm P, Opuszko M, Emde S. Using Behavioural AI to Fingerprint Large Language Models. International Conference on AI Research. 2026;7(1):34-44. doi:10.34190/icair.7.1.5580

BibTeX
@article{Bhm2026,
  author = {Böhm, Paul and Opuszko, Marek and Emde, Simon},
  title = {Using Behavioural AI to Fingerprint Large Language Models},
  journal = {International Conference on AI Research},
  year = {2026},
  volume = {7},
  number = {1},
  pages = {34--44},
  doi = {10.34190/icair.7.1.5580}
}
Kostenlos starten →