What AI can now do·Jul 22, 2025·Medicine

AI copilot cuts diagnostic errors 16% across 39,849 clinic visits in Kenya

OpenAI / Penda Health

16%Relative reduction in diagnostic errors

In Penda Health's Nairobi primary-care clinics, clinicians randomly given a GPT-4o-based 'AI Consult' made 16% fewer diagnostic errors and 13% fewer treatment errors across 39,849 visits, judged from 5,666 randomly selected charts rated by 108 physicians.

How far it has travelled

From a demonstration to real life, in five steps. Empty steps are the point: a demonstration is not a deployment.

Demonstratedshown once, by its developer or a researcher
yes · Jul 22, 2025 · Half of clinicians randomly assigned AI Consult (Jan 30-Apr 17, 2025): diagnostic errors -16%, treatment errors -13%, history-taking errors -32%; model GPT-4o. source
Reproducedshown again by someone else
partial · Jul 22, 2025 · A separate randomized controlled trial with PATH was described as ongoing; no independent replication published per the fetched page. source
Accessibleavailable to the public or to customers
partial · Jul 22, 2025 · Deployed in Penda's 16 clinics; not a general product; OpenAI provided funding and consulted on development. source
Reliableworks most of the time on a defined task
partial · Jul 22, 2025 · Error reductions are on physician-rated charts; patient-reported outcomes at eight days showed no significant difference (3.8% vs 4.3% not feeling better); peer-review status not stated. source
Real consequenceshas changed something in the world, for good or ill
yes · Jul 22, 2025 · Projected about 22,000 diagnostic and 29,000 treatment errors averted per year at Penda; 7 vs 5 safety reports with no harm attributed to the tool; clinicians asked for better localization and speed. source

What this could change for you

In clinics short of doctors, an AI second opinion already changes what patients are diagnosed with, though it has not yet been shown to make them feel better.

Sources

primary source. Evidence confidence: medium.

A milestone is a dated, sourced result that moved what AI can do; a stage is upgraded only when a source shows it (a reproduction, a public release, production evidence, a documented consequence). “Unknown” is the honest default.