Gender attribution under workplace role-and-register prompts in contemporary large language models

Abstract Large language models (LLMs) can add demographic identities when workplace descriptions leave identity unspecified. We audited nine live model configurations using 20 experimental and five control phrases, with 50 stateless generations per model–phrase cell at temperature 1.0 (11,250 responses). The experimental conditions bundled role content with communicative register: authority/resource-control phrases were direct or directive, whereas coordination/assistance phrases were reporting, deferential, or permission-seeking. The design therefore estimates a role-and-register contrast, not a causal effect of occupational status alone. Both conditions were majority female: full-grid attribution was 63.58% female, 35.89% male, 0.47% nonbinary, and 0.07% unparsed in the authority/directive condition, versus 92.62%, 6.04%, 1.20%, and 0.13%, respectively, in the support/deferential condition. Model risk differences ranged from − 12.2 to + 93.6 percentage points. A Paule–Mandel random-effects synthesis showed extreme heterogeneity (I2 = 98.6% on the log-odds scale; 95% prediction interval for the odds ratio, 0.014–63,659), so no stable common model effect is claimed. A cross-classified block bootstrap retaining 50-response cells estimated an aggregate male-attribution gap of 29.9 points (95% CI, 6.0–55.6). Name entropy diagnostics documented substantial within-cell repetition. Control mismatches were predominantly explicit model outputs rather than parser failures; excluding models below 80% control agreement did not reverse the aggregate gap. Name-based race inference is confined to a reliability appendix. The results identify heterogeneous demographic defaults under this prompt bundle and motivate crossed role-by-register experiments.