AI That Thinks Like Us? Researchers Develop New Model to Predict Human Behavior
Self-driving cars must anticipate whether a pedestrian will cross the street to avoid accidents. Investment algorithms need to predict how human traders will respond to market news before making decisions. These scenarios require machines to do more than calculations—they need to grasp human behavior. Current general-purpose AI models like GPT or Llama aren't built for this kind of insight.
Introducing Be.FM (Behavioral Foundation Model), an AI system developed collaboratively by researchers at the University of Michigan, Stanford University, and MobLab. Be.FM is among the first AI models specifically trained to predict, simulate, and reason about human actions. The research is available on the SSRN preprint server.
Specialized Training on Behavioral Science Data
Unlike traditional models trained on broad, generic text like Wikipedia, Be.FM uses a dataset curated from behavioral science research. This includes controlled experiments, surveys, and thousands of academic studies, encompassing data from over 68,000 experimental subjects and around 20,000 survey respondents.
"We built a behavioral dataset to help the model reason about why people act the way they do," explains Yutong Xie, lead author and doctoral student at the University of Michigan. This focused data helps Be.FM overcome the limitations of general AI models, which often miss minority behaviors or misunderstand complex social cues.
Four Key Capabilities of Be.FM
Be.FM exhibits several emerging skills that were not explicitly programmed. These fall into four primary application areas:
- Predicting Human Behavior: Be.FM can forecast decisions and cooperation levels in real-world situations. For example, it can predict which investment options a group is likely to prefer and how many individuals may take risks. This ability could improve economic modeling, product testing, and policy analysis by simulating group behaviors before actual implementation.
- Deducing Psychological Traits and Demographics: By analyzing behavior or background data, Be.FM can infer traits like extroversion or agreeableness, as well as demographic details such as age. This allows for better user segmentation, personalized interventions, and informed product design.
- Detecting Contextual Drivers in Behavior: Human actions can shift due to timing, social norms, or environmental factors. Be.FM identifies what might cause changes, such as app design updates or seasonal trends, by analyzing patterns across different scenarios. This insight is valuable for researchers, designers, and policymakers.
- Supporting Behavioral Science Research: Built on a large language model architecture, Be.FM can generate research ideas, summarize literature, and solve applied behavioral economics problems. It serves as a tool for brainstorming hypotheses, planning studies, or simulating scenarios before real-world testing.
Performance and Limitations
Across these areas, Be.FM outperforms commercial and open-source models like GPT-4o and Llama in matching human behavior, especially in personality prediction and scenario simulation. Its predictions align more closely with real-world population patterns.
However, Be.FM's capabilities beyond these four domains have yet to be tested. It is not currently designed to forecast large-scale political events or outcomes like elections or peace agreements. The research team is actively working to expand its domain coverage.
According to Qiaozhu Mei, professor of information at the University of Michigan and corresponding author of the study: "Behavior in health, education, even geopolitics—the goal is to make Be.FM useful wherever people make decisions."
Access and Collaboration
The Be.FM models are available upon request. The developers encourage researchers and practitioners to apply the model in their work and provide feedback to refine its capabilities.
For more information, see the original study: Yutong Xie et al, Be.FM: Open Foundation Models for Human Behavior, SSRN (2025).
Your membership also unlocks:
