The rule. Synthetic users can shape a study before it runs; only real participants can be evidence inside it.
Why. Arora, Chakraborty, and Nishimura (2025) found that LLM simulated respondents reproduce the direction of human effects while missing their magnitude and their variability, which is exactly where effect sizes and edge cases live. The strongest result in the literature is also narrower than the category's marketing: Park et al. (2024) built agents from two hour interviews with 1,052 people, and those agents matched their own participant's General Social Survey answers at 83 percent of that person's two week test retest consistency, against 74 percent for a demographics only model. That is a decent replica of someone you already interviewed, not a stand in for someone you have not. Rosala and Moran (2024) name the failure mode teams actually hit: synthetic users are agreeable, greet new concepts as a game changer, and go vague under probing.
Seen in the wild. Synthetic Users, one of the better known products in the category, headlines its site "User research, without" and then concedes in its own FAQ that real user research "stays essential for validation and edge-case work" (verified as of August 31, 2026).