Google DeepMind paper finds AI models cannot make truly random choices on their own
A new Google DeepMind paper finds that AI models cannot make truly random choices on their own, but larger models can select correctly when a program first supplies random numbers. The study had Gemini, Qwen3 and OLMo each sample 1024 times from simple distributions, and all statistical tests for matching the target probabilities failed; adjusting temperature or top-p did not fix this either. In a security game where a true randomness check would let an attacker win 75% of the time, Gemini's choices let the attacker's win rate reach 96.6% to 99.9%. The authors recommend that when agents need randomness, it should be generated in code and passed in, rather than letting the model randomize on its own.
Why read this
AI score 67
The paper used statistical tests and attack-defense games to quantify that models cannot autonomously produce randomness, and offered an engineering practice of delegating random number generation to code.
Topics
Body
New Google DeepMind paper finds that AI models cannot make truly random choices on their own, but larger models pick correctly when a computer program gives them a random number first.
Agents often need to pick actions at random with set odds, like checking servers in an unpredictable order. LLMs say they will pick uniformly, then keep landing on favorites like 7, 42, or answer C.
They asked Gemini, Qwen3, and OLMo models to sample from simple distributions 1024 times each. They also tried other temperature and sampling settings, and gave models a random number to convert.
Every reported statistical test for matching the target odds failed, and changing temperature or top-p did not fix it. In a security game where truly random checks let attackers win 75% of the time, Gemini's picks let them win 96.6% to 99.9%.
If your agent needs randomness, generate it in code and pass it in, instead of asking the model to be random.