CNOOC Discovers Over 100 Million Tons of Oil Equivalent at Qinhuangdao 29-6 Oilfield in Bohai Sea, Boosting China’s Offshore Energy Reserves

The study, reported by TechXplore, evaluated models such as ChatGPT-4o and Claude-Sonnet-4 using a classic game theory experiment called the “Keynesian beauty contest.” In this experiment, participants play a “Guess the Number” game where each player selects a number between zero and 100. The winner is the player whose number is closest to half the average of all chosen numbers. Success requires not just picking a favorite number, but predicting how others will behave and reasoning one step further.

Researchers provided the AI models with descriptions of human opponents, ranging from first-year undergraduates to experienced game theorists. While the models demonstrated some strategic thinking by adjusting guesses based on the perceived skill of their opponents, they consistently made a critical misjudgment: they assumed humans would use perfect logic and reason deeply. As a result, the AI often “played too smart,” choosing mathematically optimal numbers near zero, which failed to reflect the simpler or less rational choices real humans tend to make.

The findings have important implications as AI becomes more integrated into economic forecasting, business negotiations, and decision-making applications. If AI systems overestimate human rationality, their predictions and strategies may be inaccurate or suboptimal. Researchers argue that for AI to be truly effective, it must model not only how humans should think, but how they actually think in practice, accounting for biases, heuristics, and imperfect reasoning.