Artificial Intelligence’s (AI) ability to predict human decision-making has been the subject of a new study, revealing a significant predictability gap between AI and humans. The study, conducted by researchers from Princeton University and the LOEWE project WhiteBox at TU Darmstadt, aimed to understand how humans make suboptimal decisions in risky gambles.
Traditionally, economists, psychologists, and cognitive scientists have examined human decision-making by studying people’s risky choices in the laboratory. These choices involve alternatives with certain probabilities and payoffs. Surprisingly, the findings consistently showed that individuals deviated from the mathematically optimum choice, resulting in financial losses.
The pioneering work of Daniel Kahneman and Amos Tversky shed light on human decision-making through Cumulative Prospect Theory and earned them the Nobel Prize in Economic Sciences in 2002. However, many anomalies and contexts still exist where predicting decisions accurately remains a challenge. Understanding and explaining why individuals deviate from the optimal choice is a complex task, with different theories proposing cognitive shortcuts, known as heuristics.
To shed light on this issue, a research team from Princeton University utilized artificial intelligence to improve the understanding of human decisions in risky gambles. They collected a vast dataset comprising over 13,000 bets and trained deep neural networks to predict human decision-making. Interestingly, the neural networks that operated with fewer theoretical assumptions achieved the best accuracy in predicting gambling decisions.
Building on this research, the researchers from the LOEWE project WhiteBox at TU Darmstadt examined the predictions resulting from combining various machine learning models with different decision datasets. They discovered substantial differences in the prediction of human decisions, with some neural networks accurately predicting decisions from the recent study dataset but failing to do so in smaller psychological experiments. It became evident that biases in datasets could lead to interaction effects between models and datasets, hindering the transferability of findings.
Drawing from these observations, the researchers developed a cognitive generative model that quantitatively explains the disparities between actual decisions in datasets and the predictions made by AI models. Professor Constantin Rothkopf from the Centre for Cognitive Science at TU emphasized that while neural networks may excel at predicting certain datasets, it doesn’t guarantee their accuracy for other human gambles or everyday decisions. The study underscores the fact that automating cognitive science with artificial intelligence requires a thoughtful combination of theoretical reasoning, machine learning, and data analysis.
In conclusion, the study highlights the existing predictability gap between AI and human decision-making. While AI and deep neural networks may outperform traditional proposals by human theorists in specific datasets, their predictions cannot be easily generalized to different datasets or more naturalistic decision-making scenarios. This research contributes to the ongoing scientific interest in understanding and explaining suboptimal human decisions, emphasizing the need for a multifaceted approach combining theoretical frameworks, machine learning techniques, and comprehensive data analysis.