An AI persona is not a user. A study just measured it
A study pitted LLMs against thousands of real users: they fail to simulate their preferences. Where AI truly helps your research, where it lies to you, and the line to draw. For designers, PMs and UX researchers.
You have probably seen the shortcut, or tried it. Instead of recruiting people and talking to them, you ask an AI to play your typical user. "You are a 35 year old PM in a scale-up, here are two mockups, which do you prefer?". The answer arrives in three seconds, clean, argued, free. The temptation is huge: user research is slow and costly, and here is a machine that promises to do it for you. A study just tested that promise against thousands of real people. The verdict is clear, and not where you would expect. The machine does not get it wrong because it is dumb. It gets it wrong because it always hands you the plausible answer, and your users are not plausible. Why it matters (designer, PM, UX researcher): an AI that "plays" your user hands you a credible average, not a real…