Scaling experiments with simulated participants could make some discoveries quicker and less costly. At the same time, relying on models that generate plausible answers without the underlying reasoning risks producing output that looks confident but hides shallow understanding. For scientists trying to map how minds work, this creates a tension between productive throughput and the ability to explain why a result occurred. How we validate these surrogates, and how we integrate them with real human data, will determine whether they advance scientific knowledge or merely multiply suggestive findings.
This topic touches on core goals of the IFHP: growing human potential equitably and reliably. Questions about who benefits from model-driven research, how inclusive simulations are, and what counts as genuine explanation will shape the next phase of cognitive science. Follow the full article to see how researchers propose to test, constrain, and use AI Surrogates so that increased productivity goes hand in hand with deeper understanding.
Many scientists are enthusiastic about the potentials of ‘Artificial Intelligence’ (AI) for research. We recently examined the vision of ‘AI Surrogates’ [1]: computer models [including but not limited to large language models (LLMs)] designed to simulate human participants for the purpose of generating knowledge about human cognition and behavior. Some scientists believe that AI Surrogates can improve the generalizability of cognitive science: first, by simulating diverse populations that are not readily accessible, overcoming the field’s overreliance on Western, Educated, Industrialized, Rich, Democratic (WEIRD) samples; and second, by enabling researchers to quickly and cheaply explore vast experimental design spaces, expanding the diversity of situations that can be probed experimentally.
