Artificial-intelligence systems built on large language models could push consumers toward safer, more conventional choices, potentially flattening cultural variety and personal expression, according to new research led by Columbia Business School professor Sandra Matz. Analyzing more than 110,000 real-world decisions from 1,000 people and comparing them with outputs from generic and personalized AI agents, the study finds that LLM-driven recommendations gravitate to the average, nudging users toward mainstream options and shrinking the range of exploration. The dynamic reflects how models are trained and optimized—to minimize risk and keep engagement high—rather than to surface idiosyncratic or serendipitous picks. The authors suggest developers add an “exploration mode” to widen recommendations and help preserve diversity in preferences. The findings add to a broader debate over AI’s impact on consumer choice, creativity and market differentiation as companies accelerate deployment of AI tools.
Related articles:
Recommender system
Filter bubble
Exploration–exploitation tradeoff
Echo chamber (media)
Algorithmic bias
