Ishan Anand

Large language models can now stand in for humans in surprising ways, from predicting personality types to replicating their responses in market research. Like weather forecasting, once considered impossible and now so routine we take it for granted, LLMs are in the early, unreliable-but-improving stage of simulating how populations think and respond. Teams are already using LLMs as synthetic survey respondents for concept testing, UX exploration, and early market validation. In the past year, the field has gotten both more promising and more tricky. The real question is no longer "can LLMs simulate people?", but whether the simulation is validated for the decision you want to make. New methods show that how you ask an LLM matters as much as which model you use and can dramatically improve fidelity to real human responses. Meanwhile validation studies show accuracy can mask subgroup distortion and that seemingly minor choices can reshape the simulated population entirely. This talk gives entrepreneurs, engineers, and PMs an overview of the techniques and a framework for validating synthetic respondents before making decisions. Even if you never build a synthetic persona, this is one of the richest windows into LLM behavior under the hood and these lessons apply to any system where you're trusting an LLM to represent something about the real world.