A landmark study published in *Nature*, led by Harvard psychology researcher Ashwini Ashokkumar alongside Luke Hewitt, Isaias Ghezae, and Robb Willer, demonstrates that models like GPT-4 can forecast the outcomes of social science experiments with astonishing accuracy. The team fed GPT-4 detailed descriptions of experimental "treatments" and demographic profiles, using data from 70 real-world experiments representing nearly 120,000 actual participants. The AI's predictions correlated strongly (0.85) with actual measured treatment effects, and GPT-4 predicted outcomes with a root mean squared error of just 5.3 points, outperforming human expert forecasters who scored 6.0 points. The simulation successfully modeled 476 different randomized treatments without recruiting a single living participant. However, the research carries a heavy caveat: accurate prediction is not synonymous with sentient understanding. US scholars Lisa Messeri and Molly J. Crockett have coined this phenomenon the "illusion of understanding," where LLMs generate outputs that appear deeply insightful, encouraging researchers to overestimate the machine's grasp of human nuance.