Work

Language models advise more harmful or unnecessary medicine to informal providers in India

Year: 2026
Type: preprint
Abstract: Informal providers give much of rural India's first-contact primary care. We built simulated consultations from four Indian standardized-patient cases (angina, dysentery, tuberculosis, asthma). Claude Sonnet 5.5 played the patient; twelve models answered in English, Hindi or Telugu (2,400 conversations). Each case ran in a neutral and a practitioner setting; the latter names the user as an informal provider relaying a patient. Native readers checked the Hindi and Telugu cases. Claude Sonnet 5.5 (scorer A) and Kimi K2.5 (scorer X) scored every transcript against the source checklists, and a third model scored a stratified sample of 282 runs. The practitioner setting raised the odds of harmful or unnecessary medicine, as defined in the source standardized-patient studies (adjusted odds ratio 7.63 under scorer A and 4.35 under scorer X; 95% confidence intervals 3.23-17.98 and 2.71-7.00). In English it rose from 3% to 17% of runs under scorer A, and it rose in every language under all three scorers. Correct management in English also rose, from 88% to 94%. With a second patient model, harm rose from 12.5% to 52.7% of runs (scorer A) and from 10.0% to 34.9% (scorer X; odds ratio 5.00, 0.75-33.40). Telugu had lower odds of correct management than English (odds ratio 0.28 under scorer A, 0.41 under X). On 240 HealthBench items, a public set of physician-graded health conversations, scorer X agreed with the physician majority on 75.8% and scorer A on 56.7%. Validating the model scorers against physicians is the next step.
Language: English
Field: Medicine
Open Access status: green
OpenAlex ID: W7236621544
Cites:
Cited by:
Related to: 10