Language models simulating social-media users were more faithful to individual people when instructed to respond intuitively and immediately, according to a small study. Asking the models to analyze their answers produced worse matches, challenging the assumption that more visible reasoning always helps.
Researchers profiled eight Serbian participants through questionnaires, interviews and written self-descriptions, then recorded their reactions to 68 posts. Four language models predicted those reactions under five prompt conditions. Attitudes provided much more useful profile information than demographic backstories. The intuitive condition performed best even on topics that the original questionnaire had not covered.
AI simulations still compressed differences between people: under the best setup, variation was flattened to three times the human level, compared with seven times under weaker conditions. The sample of eight is far too small to justify broad claims about populations or elections. The result is a prompt-design finding that needs replication, and it also highlights a misuse risk: more convincing synthetic profiles could manipulate perceptions of public opinion.