A new field, AI Behavioral Science, is rapidly emerging at the intersection of artificial intelligence and behavioral science. This paper, co-authored by researchers from leading institutions including Stanford, MIT, and the University of Michigan (with affiliations noted from organizations like MobLab), introduces the core questions and three reinforcing themes that define this crucial area of study.
As complex AI systems are increasingly shaping human experiences, from influencing decisions on social media to automating tasks in the labor market, understanding and guiding these interactions is vital.
The Three Reinforcing Pillars
The field is structured around a mutually reinforcing cycle of themes:
Assess AI Behavior
Behavioral science methods are used to study and design AI. Since AI's code is complex and often proprietary, the goal is to evaluate AI based on its observable behavior.
Core Question: What objective or motive is an AI agent implicitly trying to achieve in different scenarios with humans and other AI?
Method: Applying diagnostic tools like Turing Tests and "revealed preference" analysis (inferring what objective an agent acts as if it is maximizing) to predict AI behavior.
Use AI for Behavioral Science
AI's enhanced computational and processing capabilities are leveraged to transform social and behavioral research.
Core Question: How can AI be used to better understand, model, and analyze complex human decision-making processes and behaviors?
Method: Simulating individuals, groups, or entire social systems at new scales; analyzing massive amounts of qualitative data (text, video); and creating digital proxies of humans to conduct rich social and behavioral experiments in silico.
Understand AI-Human Interactions
This theme focuses on the complex and layered new systems created when humans and AI interact.
Core Question: How does AI influence human behavior, and how do both AI and humans change their beliefs and behaviors when interacting with each other?
Significance: Understanding how AI systems change human knowledge, behavior, and the shifting economy (e.g., jobs being augmented or displaced) is vital for steering these interactions toward positive social outcomes.
The convergence of these three areas is essential for developing methods to better predict how AI will behave, guide its design for human and social welfare, and predict and manage the societal impact of this technological revolution.