Back

New Model, Old Risks? Sociodemographic Bias and Adversarial Hallucinations Vulnerability in GPT-5

Omar, M.; Agbareia, R.; Apakama, D. U.; Horowitz, C. R.; Freeman, R.; Charney, A.; Nadkarni, G.; Klang, E.

2025-09-21 health informatics
10.1101/2025.09.19.25336180 medRxiv
Show abstract

Plain summaryExtending our validated benchmarking work, GPT-5 showed no improvement in sociodemographic-linked decision variation compared with GPT-4o and seemed to be worse on several endpoints. We re-tested GPT-5 with a fixed pipeline: 500 physician-validated emergency vignettes, each replayed across 32 sociodemographic labels plus an unlabeled control, answering the same four questions (triage, further testing, treatment level, and need for mental-health assessment). This design holds clinical content constant to isolate the effect of the label. GPT-5 reproduced subgroup-linked variation, with higher assigned urgency and less advanced testing for several historically marginalized and intersectional groups. Notably, several LGBTQIA+ labels were flagged for mental-health screening in 100% of cases, versus ~41-73% for comparable groups with GPT-4o. Additionally, in an adversarial re-run that inserted one fabricated medical detail into otherwise standard clinical cases, GPT-5 adopted or elaborated on the fabrication in 65% of runs (vs 53% for GPT-4o). A single mitigation prompt reduced this to 7.67%.

Published in npj Digital Medicine (predicted rank #1) · training set

Matching journals

The top 2 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.