LLMs Can Better Capture Human Judgments--With the Right Prompts

In three linesSimple prompting techniques improve LLMs' ability to capture human judgments. Across 144 moral scenarios (US) and 38 moral beliefs (32 countries), asking models to report standard deviations and response proportions better recovers the full distribution of human responses. LLMs also track human confusion ratings, though their error calibration remains poor.

Summary generated by Claude — human-verified