Do AI models think like humans when making risky bets?

Researchers tested 28 large language models on the St. Petersburg paradox—a classic puzzle where expected payoff is infinite but humans reasonably bid low. Models did produce human-like bids, but probing with variants revealed they weren't using human reasoning: when games changed (repeated play, different payoffs), models shifted to calculating rationally rather than sticking to human-like patterns. This matters because AI systems deployed for high-stakes decisions might look aligned with human judgment while operating on alien principles underneath.