AI Overestimates Human Intelligence and Can Make Wrong Decisions
Scientists have discovered that current AI models tend to overestimate the rationality of their human opponents, whether they are first-year university students or experienced scientists, in strategic thinking games and other tasks. This could be one of the greatest risks when delegating complex strategies to artificial systems.
Researchers from the Higher School of Economics in Russia found in a study published in the Journal of Economic Behavior & Organization that large language models (LLMs), including ChatGPT and Claude, tend to overestimate the rationality of their human opponents in classic strategic games, such as the so-called 'Keynesian beauty contest'. This overestimation leads AI models to choose more sophisticated strategies than humans actually use, and in many games, this results in defeat.
In the 1930s, British economist John Maynard Keynes developed the theoretical concept of a metaphorical beauty contest. According to a press release, the classic example involves asking newspaper readers to select the six most attractive faces from a set of 100 photographs. The prize goes to the participant whose choices are closest to the most popular selection, or the average of the selections of the rest of the participants. In most cases, people tend to select the photos they find most attractive by a subjective choice. However, this usually leads to defeat, because the real objective is to predict which faces will be most attractive to the majority of respondents. Thus, a rational strategy would be to base choices on other people's perceptions of beauty.
The scientific work faced LLMs against populations of different profiles: from first-year university students to attendees at game theory conferences. Each model was explained the rules of the game and asked to choose and justify their reasoning. The authors observed that, although LLMs adapt their choices according to the expected sophistication of the adversary, their starting point is usually more 'reasonable' or closer to the theoretical equilibrium than that of most humans.
According to the researchers, the models use abstract and balanced reasoning patterns that frequently appear in texts and manuals: that is, they learn the idealized version of a player and thus see all their opponents. In this way, they do not fully capture the cognitive limitations or real biases that shape everyday human decisions. In practice, assuming that a human opponent will reach a reasoning superior to the real one can lead to a theoretically correct decision, but misaligned with the observed behavior of the general population, which is precisely the game's objective.
The finding has implications in finance and markets: for example, the 'beauty contest' theory is used to explain how economic agents anticipate the expectations of others. An AI model that assumes too much rationality could recommend strategies that clash with market reality and generate losses or erroneous decisions.
Researchers from the Higher School of Economics in Russia found in a study published in the Journal of Economic Behavior & Organization that large language models (LLMs), including ChatGPT and Claude, tend to overestimate the rationality of their human opponents in classic strategic games, such as the so-called 'Keynesian beauty contest'. This overestimation leads AI models to choose more sophisticated strategies than humans actually use, and in many games, this results in defeat.
In the 1930s, British economist John Maynard Keynes developed the theoretical concept of a metaphorical beauty contest. According to a press release, the classic example involves asking newspaper readers to select the six most attractive faces from a set of 100 photographs. The prize goes to the participant whose choices are closest to the most popular selection, or the average of the selections of the rest of the participants. In most cases, people tend to select the photos they find most attractive by a subjective choice. However, this usually leads to defeat, because the real objective is to predict which faces will be most attractive to the majority of respondents. Thus, a rational strategy would be to base choices on other people's perceptions of beauty.
The scientific work faced LLMs against populations of different profiles: from first-year university students to attendees at game theory conferences. Each model was explained the rules of the game and asked to choose and justify their reasoning. The authors observed that, although LLMs adapt their choices according to the expected sophistication of the adversary, their starting point is usually more 'reasonable' or closer to the theoretical equilibrium than that of most humans.
According to the researchers, the models use abstract and balanced reasoning patterns that frequently appear in texts and manuals: that is, they learn the idealized version of a player and thus see all their opponents. In this way, they do not fully capture the cognitive limitations or real biases that shape everyday human decisions. In practice, assuming that a human opponent will reach a reasoning superior to the real one can lead to a theoretically correct decision, but misaligned with the observed behavior of the general population, which is precisely the game's objective.
The finding has implications in finance and markets: for example, the 'beauty contest' theory is used to explain how economic agents anticipate the expectations of others. An AI model that assumes too much rationality could recommend strategies that clash with market reality and generate losses or erroneous decisions.
Additional Information (Preview)
Artificial Intelligence (AI) has made significant progress over the last few decades, evolving from simple programs capable of executing specific tasks to complex systems that can learn, adapt, and make decisions based on vast amounts of data. These advancements have been made possible through a combination of more sophisticated algorithms, an exponential increase in computing power, and access to large datasets. However, as AI becomes increasingly integrated into our lives, new questions and challenges arise about how it interacts with and understands human behavior. One of the most fascinating fields of study has been the interaction between AI and humans in strategic decision-making contexts. This not only has applications in games and competitions but also in critical areas such as negotiation, financial planning, and security. The ability of AI to...
This is a preview of additional information. Register to see the full text.
Source: Levante-EMV • All content and images belong to their respective owners.
