Auto-Psych Runs Full Psychology Experiments With AI Agents

Stanford SAIL's post points to auto-psych, where an inner loop generates, fits, and critiques a probabilistic model of some piece of human behavior, and an outer loop decides which experiment would distinguish between competing candidate theories, then actually runs that experiment with paid participants recruited online and feeds the results back into the loop.

The authors report that across three independent sequences of human experiments, the theories this system arrived at fit the collected data better than theories drawn from the existing scientific literature on the same topics. That result describes fit to data within the specific behaviors studied, not that the system's theories are broadly correct or have been independently replicated.