AI models overestimate smartness of people: Study
The rapid advancement of artificial intelligence (AI) has led to the development of sophisticated models that can simulate human-like conversations, play complex games, and even exhibit creative behaviors. However, a recent study by scientists at HSE University has revealed a fascinating flaw in these models: they tend to overestimate the smartness of people. This phenomenon was observed in popular AI models, including ChatGPT and Claude, which often play “too smart” and lose games due to their assumption of a higher level of logic in humans than is actually present.
To investigate this phenomenon, the researchers employed the Keynesian beauty contest, a classic game that requires strategic thinking and social reasoning. The game is simple: each player is asked to choose a number between 0 and 100, and the winner is the one who selects a number closest to two-thirds of the average number chosen by all players. The twist is that players must think about what others will think, creating a complex web of strategic reasoning.
The researchers tested the AI models on this game, pitting them against human players. The results were striking: while the AI models were able to play the game with remarkable skill, they consistently overestimated the smartness of their human opponents. This led to a phenomenon known as “over-strategizing,” where the AI models would play “too smart” and lose the game due to their assumption of a higher level of logic in humans than was actually present.
For example, in one iteration of the game, an AI model might choose a number based on its assumption that human players will choose numbers based on a complex analysis of the game’s dynamics. However, if the human players are not as sophisticated in their thinking, the AI model’s strategy will backfire, and it will lose the game. This pattern was observed repeatedly in the study, with the AI models consistently overestimating the smartness of human players and paying the price for it.
The study’s findings have significant implications for the development of AI models. If these models are to be used in real-world applications, such as decision-making or negotiation, they must be able to accurately assess the cognitive abilities of their human counterparts. Otherwise, they risk over-strategizing and making suboptimal decisions.
The researchers suggest that the overestimation of human smartness by AI models is due to their training data, which often consists of idealized scenarios and rational actors. In reality, humans are often driven by biases, emotions, and cognitive limitations, which can lead to suboptimal decision-making. The AI models, lacking this nuance, assume that humans will behave in a more rational and sophisticated manner than is actually the case.
To address this issue, the researchers propose that AI models be trained on more diverse and realistic data sets, which reflect the complexities and limitations of human cognition. This could involve incorporating data from behavioral economics, psychology, and sociology, which can provide a more nuanced understanding of human decision-making.
In conclusion, the study by HSE University scientists highlights a critical flaw in current AI models: their tendency to overestimate the smartness of people. This phenomenon has significant implications for the development of AI models and their potential applications in real-world scenarios. By acknowledging and addressing this issue, researchers can create more sophisticated and human-like AI models that are better equipped to interact with and understand humans.
As the field of AI continues to evolve, it is essential to recognize the limitations and biases of current models. The study’s findings serve as a reminder that AI models are only as good as their training data and that they must be designed to account for the complexities and nuances of human cognition. By doing so, we can create AI models that are more effective, more efficient, and more human-like in their interactions with us.
News Source: https://www.sciencedirect.com/science/article/abs/pii/S0167268125004470
About the Author
