A study led by Columbia Business School professor Sandra Matz suggests that large language models (LLMs) used in artificial intelligence may lead to more predictable human behavior by recommending common or average choices. The research indicates that because AI models are trained to predict the most likely next word or event, they tend to provide normative outputs that can homogenize user decisions.
The study analyzed the behavior of LLMs to determine how they influence personal preferences. Matz, a computational social scientist, focused on how these systems interact with individual decision-making processes across different topics and psychological affinities.
To conduct the research, Matz and her co-authors analyzed more than 110,000 real-world decisions made by 1,000 people. They compared these human choices to recommendations provided by both generic and personalized AI agents. The team also utilized data from the myPersonality project, a Facebook application that administered personality tests to users who shared their profiles for research purposes.
The researchers noted that this behavior is a result of how AI is currently programmed rather than an inherent limitation of the technology. According to Matz, the reliance on these models could narrow the range of what individuals explore, potentially leading to a collapse of culture into a single set of preferences. The study suggests that if these patterns continue, the distinctiveness and "quirky" behaviors that characterize individual human experience could be reduced.
To address these findings, Matz recommended that technology developers introduce an "exploration mode" for AI agents. This feature would allow users to opt for more unexpected or less conventional recommendations, rather than the standard normative outputs. No specific timeline for the implementation of such features by tech companies was reported.