Synthetic respondents validity holds up in categories a model already knows, which is exactly why the pressure to use them everywhere is real. For an Insights Lead, the harder question is where that validity actually stops.

A Columbia Business School team trained digital twins on more than 500 questions per person, then checked each twin against the real person it modeled. The correlation was roughly 0.2, where 1.0 would be a perfect match, and the twins missed the attraction effect that drives real pricing and concept trade-offs (Peng et al., 2025, revised 2026).

The full paper adds seven more studies and EPIC’s own conjoint benchmark, turning them into a rule for where AI can accelerate research and where a decision still needs real human choice data.