A new study co-authored by Columbia Business School professor Sandra Matz indicates that large language models (LLMs) may steer users toward common, risk-averse choices and decrease individual exploration. Researchers compared human decisions to choices made by both generic and personalized artificial intelligence agents.

Matz, a computational social scientist with a background in psychology and computer science, authored the study on how large language models affect human decision-making. She and her co-authors analyzed more than 110,000 real-world decisions made by 1,000 individuals, using data from a Facebook application known as the myPersonality project. This project conducted personality tests on users who volunteered their Facebook profiles for research.

Matz stated that LLMs inherently promote average outcomes. "LLMs predict the most likely next word in a sentence or event in a sequence, and by definition, that's average," Matz said. "It homogenizes decisions, and we all get the same output." She noted, "It tells you what the most likely thing to appear is if you ask it for a movie recommendation or what color to paint your wall."

According to Matz, these AI models avoid risk due to their training. "AI hates risk because we train it that way," she said. This tendency can cause LLMs to keep users within familiar confines. She explained, "It wants to keep you on the platform, so it shows you what you already like and not stuff on the outskirts of what you do."

Matz's study found that "LLM agents nudge behavior toward more normative options and narrow the range of what individuals explore." She suggested that tech developers should integrate an "exploration mode" feature for users. Matz believes this feature "would help ensure we prevent ourselves as individuals from becoming boring, and making sure culture doesn't collapse into a single set of preferences."

Why It Matters

The findings from this study suggest that the widespread adoption of large language models could influence individual decision-making by promoting conformity. If users consistently receive recommendations and information based on the most probable or average outcomes, it could limit their exposure to diverse ideas, products, and experiences. This potential effect has broader implications for cultural evolution and individual expression if the range of explored options narrows across a large user base.