This post is also available in:
עברית (Hebrew)
Artificial intelligence has made impressive strides in recent years, but when it comes to understanding how people behave in real-world settings, most models still fall short. A new AI system, known as Be.FM (Behavioral Foundation Model), is looking to close that gap.
Developed by researchers from the University of Michigan, Stanford University, and MobLab, Be.FM is among the first AI models purpose-built to predict and reason about human behavior across different situations. Unlike general-purpose models such as GPT-4o or LLaMA, Be.FM is trained on behavioral science data rather than broad text sources like Wikipedia or news articles.
This data includes controlled experiments, academic surveys, and findings from published behavioral studies—covering over 68,000 experimental subjects and 20,000 survey responses, according to TechXplore. This tailored training allows the model to capture a wider range of human decisions and social cues that traditional AIs often overlook.
Be.FM’s strength lies in four key areas. First, it can forecast human decision-making in practical scenarios. For instance, when presented with investment choices, the model can estimate how many people might take risks, cooperate, or opt for safer options—useful for economic planning, product testing, or policy evaluation.
Second, the model can infer psychological and demographic traits based on behavioral inputs. That means it can help researchers or organizations better understand audience segments or personalize interventions without direct access to survey data.
Third, Be.FM can analyze how behavior changes in response to context—such as time of year, interface design, or shifting social norms. For example, it can help identify why users respond differently to the same app feature in different times of the year.
Lastly, the system can assist in research workflows. It supports literature reviews, hypothesis generation, and behavioral modeling, offering a practical tool for academics and professionals in behavioral economics or policy.
Initial benchmarks show that Be.FM outperforms general-purpose AI models in tasks like personality prediction and scenario-based simulation. However, its capabilities are still limited to behavioral domains and don’t extend to areas like political forecasting.
The Be.FM models are currently available upon request, with the development team encouraging broader testing and collaboration.
The research is available on the SSRN preprint server.
