How Perceived Socioeconomic Background Influences AI-Generated Health Communication
Candace Pun
13/11/27
Large language models are increasingly used as sources of health information, but it is unclear whether they tailor communication to individual users or rely on assumptions about a user’s inferred socioeconomic or linguistic background, potentially creating differences in communication quality or accessibility across patient groups.
We tested three commercial AI models (GPT-4o, Gemini 3.5 Flash, Claude Sonnet 5) using nine personas spanning a range of socioeconomic and educational contexts, including six with matched identity and query register, three with deliberately mismatched register, across three chronic conditions (hypertension, diabetes, Crohn’s disease) with three replicate generations each, yielding 243 responses. Input and output readability (Flesch-Kincaid Grade Level) and response quality were measured for each response.
All three models adjusted output complexity, but the magnitude of adaptation was weak, with responses compressing toward an intermediate readability range of grades 9-11 regardless of how simply or elaborately a user wrote. This adaptation was uneven across socioeconomic contexts: models adjusted substantially to query-complexity changes for higher socioeconomic tier personas but showed no detectable adjustment for lower socioeconomic personas, and socioeconomic tier remained associated with output readability even after accounting for input complexity. Greater response complexity was not associated with higher quality; mismatched identity-register conditions scored significantly higher on quality than matched ones, though quality metrics are exploratory and were not independently verified. A limited-English-proficiency persona received markedly more complex responses despite writing the dataset's simplest queries.
These findings suggest that current AI models adapt communication style incompletely and unevenly across socioeconomic contexts, independent of response quality. Patient-facing AI should prioritize a user’s demonstrated communication needs over cues inferred from occupational or socioeconomic contexts.