AI agents may make human choices more predictable, study finds

AI agents that act on a user’s behalf made choices more predictable and less varied over time, a Columbia Business School study found in a review of 110,000 real-world decisions from 1,000 U.S. social media users. The work, led by Sandra Matz, a Columbia Business School professor and the Lulu Chow Wang Professor of Business Management, compared unaided decisions with generic AI agents and personalized AI agents.

The pattern was split in two. Generic AI made different people’s choices more similar, while personalized AI preserved some individuality across users but narrowed exploration within each person. In Columbia Business School’s Oct. 9, 2025 research brief, Matz and co-authors C. Blaine Horton and Sofie Goethals argued that the systems “play it safe within a user’s preferences,” nudging people toward more conventional options rather than broader exploration.

That finding raised a sharper question than convenience alone. Columbia said the research touched human individuality, creativity, cultural output and the risk that AI systems could flatten taste, reduce serendipity and encourage more normative choices. The concern is not only that two people may begin to sound alike, but that one person may become easier to predict with each passing use, as an AI agent learns which options are safest and most familiar.

The Columbia results fit with warnings from Harvard Business School, which in a July 2024 paper said artificial intelligence can reduce the range of people’s expression and choices and shape their “possibility spaces.” A separate Harvard Business Review article published in January and February 2025 said many people resist AI because they see it as opaque, emotionless, rigid and independent, with fear of job losses and data misuse also weighing on adoption.

Matz has also described AI-driven psychological targeting as a way digital footprints can be used to influence choices, underscoring a broader worry that the same systems designed to save time could also narrow judgment. Her public framing points to a deeper trade-off now visible in the data: efficiency may come at the cost of variety, and the more an agent learns a person’s habits, the more it may steer that person toward the familiar.