Tell an AI to stop overthinking, and it gets better at pretending to be you.
Researchers profiled eight Serbian participants using a questionnaire, a deep interview, and a written self-presentation, then recorded how each person actually reacted to 68 real social media posts. Four language models were then asked to predict those same reactions under five different prompting setups, ranging from bare demographics to rich attitudinal profiles paired with different instruction styles. Profiles built from attitudes and opinions beat plain demographic backstories by a wide margin. The biggest single gain, though, came from telling the models to answer immediately and intuitively rather than reason it out - that instruction cut the models' tendency to flatten individual differences from seven times the human level down to three.
That prompting style didn't just help on topics participants had already discussed in their questionnaire - it also produced the best results on posts about subjects nobody was profiled on, beating a simple crowd-average baseline. That suggests intuition-prompted agents could work as general-purpose stand-ins for real users, not narrow specialists trained on a fixed topic list. Given that platforms increasingly test policy changes on synthetic user panels, and that the same technique could just as easily help manufacture convincing fake opinion ahead of an election, how well these agents mimic real people cuts both ways.
The paper's own framing captures the irony: making AI simulate people better meant asking it to think less, not more.