How Many Guesses Does It Take to Crack an AI’s “Random” Number?

Quick take

The way AI models generate numbers labeled as “random” is far from truly random. By analyzing the exact probability outputs, or logits, from both base and instruction-tuned language models, researchers can estimate how many guesses it takes to predict these so-called random numbers. This involves converting the logits into a measure called bits of entropy, which directly relates to how predictable the output is. The lower the entropy, the fewer guesses needed to crack the number.

Why it matters

The label “random” in AI outputs can mislead operators about the unpredictability of model-generated data. If the AI’s “random” numbers have less entropy than expected, attackers or competitors can guess values with fewer attempts. This weakens any security or fairness assumptions based on the randomness of these outputs. Builders and users relying on AI for tasks involving randomness should reassess how secure or unbiased their systems truly are. This pressure might shift design decisions toward incorporating true randomness or cryptographically secure methods when unpredictability matters.

AI Quick Briefs Editorial Desk