Answers are grounded in this paper and cite their sources.
Metamorphic audit reveals systematic under-diagnosis of treatable diseases in low-income settings across frontier models, highlighting persistent algorithmic biases unaffected by scale.
This preprint reports a metamorphic audit showing that large language models systematically under-call treatable cosmopolitan disease (autoimmune conditions, cancers, non-communicable disease) when a patient is framed as being in a low-income setting. This under-call rises with model scale, is not reliably corrected by medical fine-tuning, and is not resolved even by frontier-tier models (Claude Sonnet 5, GPT-5.6, Gemini 3.5), which under-call at rates matching the largest open-weight models tested. A methodological confound in metamorphic LLM auditing, using a subject model to judge its own output is separately identified, quantified, and mitigated with a deterministic matcher independently validated against clinician adjudication. This is the author's original manuscript version, currently under review at the Journal of the American Medical Informatics Association (JAMIA).
No takes yet. Share an insight, caveat, or question.
Eke et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: