The paper examines how platform policies are increasingly evaluated using artificial users, making the fidelity of simulated agents crucial while also warning that fabricated profiles could manipulate public opinion before elections. Eight Serbian participants were profiled through a questionnaire, a deep interview, and a written self‑presentation, and their reactions to 68 social‑media posts were recorded. Four language models were then asked to predict those reactions under five prompting conditions that varied profile content (demographic backstory vs. attitudinal description) and instruction style (analytical vs. intuitive).
The findings reveal that prompts containing attitudinal information dramatically outperform those with only demographic backstories. Models aligned with the supplied profiles more closely than participants aligned with their own survey answers, and consistency became unrelated to fidelity once profile information was present. Instructing models to respond "intuitively and immediately" yielded the highest fidelity, reducing the compression of individual differences from seven times the human level to three times.
This advantage persisted on posts about topics never raised in the questionnaire, where the intuitive condition outperformed a crowd baseline by a wide margin, suggesting that such prompted agents could serve as general‑purpose simulated users rather than topic‑specific specialists. The results imply that intuition‑based prompting may be better suited for certain tasks than reasoning‑based approaches, offering insights for future LLM development.
Review