Original Source
Mimicking Human Biases Improves AI Learning
New Approach to AI Learning
According to research from the University of Chicago Booth School of Business, while people often perceive AI as cold and rational, incorporating human misperception in training could make AI smarter and cheaper. Sijia Liu, a PhD student at Princeton University, Niklas Muennighoff, a PhD student at Stanford University, and Kawin Ethayarajh from Chicago Booth examined different approaches to aligning AI models after pretraining. They found that better-performing methods serendipitously reflect human biases about probabilities.
Cost-Effective AI Alignment Techniques
Training state-of-the-art language models is expensive, with a significant portion of costs stemming from alignment—the process of refining raw model capabilities into useful outputs. The researchers explain that costly online alignment methods succeed partly because they accidentally force language models to learn in a way that mirrors human psychology. This insight enabled them to create an approach that matches the quality of expensive methods at a fraction of the cost. Humans tend to systematically distort probability, overestimating rare outcomes while underestimating more common ones, a phenomenon known as probability weighting in behavioral economics.
*Source: The University of Chicago Booth School of Business (2026-05-07)*
Related Articles
📧 Daily Newsletter
Get the daily global news briefing in your inbox every morning.
It's still free.