Recent research reveals that while LLMs can replicate median survey responses to within a percentage point, they fail to capture the true diversity of human opinions. When asked to simulate 6,000 American households on inflation expectations, models like Llama-3 match the median response but place 95% of respondents within a narrow two-percentage-point window. This fundamental limitation calls into question their utility as survey replacements.

The "mode collapse" phenomenon persists despite standard remedies. Researchers tested five leading LLMs against real survey data and found that 44-70% of human respondents give answers more than 3 percentage points away from the modal reply. In contrast, LLM samples show virtually zero such variation. Census-derived personas and knowledge-cutoff instructions fail to address this core limitation, suggesting the models rely on memorized statistics from training data.

Two "unlearning" methods show promise in addressing these limitations. Gradient Ascent and Negative Preference Optimization remove memorized statistics from model weights rather than simply instructing models to ignore them. After unlearning, exact mode matches dropped from 92% to 24-37%, with 43-43% of responses moving beyond ±3 percentage points. This approach recovers a more realistic response distribution while maintaining general survey reasoning capabilities.