Crockett and Messeri build on these concerns by focusing on how AI surrogates can reinforce the same sampling biases researchers hope to correct. Models trained on large, but not necessarily diverse, datasets can reproduce assumptions tied to WEIRD populations. The tasks used in many studies often strip away interaction, history, and physical environment, which are central to how people form beliefs and make decisions. Paying attention to these limitations matters for designing experiments that speak to real-world lives.
For anyone interested in expanding human potential or making science more inclusive, the stakes are practical. Methods that overlook embodied and social aspects of cognition risk producing policies and tools that fail where they are most needed. The article invites readers to rethink how AI tools are used in behavioral research and to consider alternative approaches that preserve context, diversity, and meaningful participation. Follow the link to learn how these ideas could reshape research practices and who benefits from their results.
Using large language models (LLMs) to replace human participants suffers from fundamental fallacies: overgeneralization from Western, Educated, Industrialized, Rich, Democratic (WEIRD) samples; conflation of linguistic form with psychological content; and neglect of embodied and social dimensions of cognition [1]. In their recent article in TiCS [2], Crockett and Messeri extend this critique, arguing that ‘AI Surrogates’ perpetuate generalizability problems by entrenching WEIRD samples and decontextualized tasks [3].
