New model, old risks: sociodemographic bias and adversarial hallucinations vulnerability in GPT-5.

Omar, Mahmud; Agbareia, Reem; Apakama, Donald U; Horowitz, Carol R; Freeman, Robert; Charney, Alexander W; Nadkarni, Girish N; Klang, Eyal · NPJ Digit Med · 2026

cross_sectional · Level IV

Where this comes from

Abstract

We re-evaluated GPT-5 using our published pipelines: 500 emergency vignettes across 32 sociodemographic labels for bias, and adversarial prompts with fabricated details. GPT-5 showed no measurable improvement over GPT-4o in sociodemographic-linked decision variation, with several LGBTQIA+ groups flagged for mental-health screening in 100% of cases. Adversarial hallucination rates were higher (65% vs 53% for GPT-4o); a mitigation prompt reduced this to 7.67%.