Nature Flags Persistent Bias and Hallucination Risks in GPT-5 Medical Diagnostics
A Nature paper reports that GPT-5 still shows sociodemographic bias and remains vulnerable to adversarial hallucinations in medical-diagnosis tasks. The findings are a reminder that frontier models may be more capable, but they are not yet reliably safe for clinical use.
This study matters because it pushes back on a common assumption: that each new model generation automatically solves the last one’s safety flaws. In medical diagnosis, bias and hallucination are not abstract technical defects; they can translate into unequal recommendations and unsafe outputs.
The mention of sociodemographic bias is especially concerning because healthcare already struggles with unequal performance across populations. If a model responds differently depending on demographic context, then its deployment risks amplifying the very disparities medicine is trying to reduce.
Equally important is the adversarial hallucination vulnerability. Medical AI is often discussed as if the main challenge is training enough data, but real-world use also requires resilience to malicious or misleading inputs. A model that performs well in benchmark settings may still fail under pressure.
The takeaway is not that frontier models are useless, but that they require stronger governance, stress testing, and human oversight than many vendors currently admit. In clinical care, cutting-edge capability is only valuable if the system can remain safe when the inputs get messy.