As a cognitive scientist, my goal is to build better models of human reasoning, but our field has always been limited by two simplifications: a focus on people as isolated individuals, and a focus on very constrained behavioral tasks. The recent accumulation of technological advances provides new tools to overcome these limitations. I will present a large behavioral study of experimental diffusion of joint-action norms, as an example for how AI tools can support scalable discovery of structured models of human reasoning from open-ended natural-language interactions between people. I will discuss current limitations and open problems with this approach, and reasons for enthusiasm.