All Public Voices Are Equal, But Are Some More Equal Than Others to LLMs?

Occupation-based bias is systematic across LLMs: When identical comments were attributed to different occupations, LLMs consistently produced lower-quality summaries for comments from lower-status workers (street vendors) compared to higher-status professionals (financial analysts), preserving less meaning and using simpler language regardless of the actual comment content.

This bias matters for democratic participation: Federal agencies are using these LLMs to process public comments in rulemaking—the primary way citizens influence regulations. Systematic differential treatment means some voices are effectively being heard less clearly by the systems making policy decisions.

Race and gender effects were minimal, but occupation effects were universal: Across all eight LLMs tested, 32 identity conditions, and multiple regulatory contexts, occupation was the only demographic signal that consistently produced differential treatment, suggesting this particular bias is deeply embedded in how these models process information.

Model choice determines the degree of bias: Different LLM providers showed varying magnitudes of occupation-based differential treatment, meaning federal agencies' selection of which model to deploy directly determines how much bias will affect public comment processing.

Surface-level writing quality isn't the driver: Experimentally adding spelling and grammar errors had negligible effects on summarization quality, indicating the bias stems from how models interpret occupational status itself rather than assumptions about writing ability.