Large language models could help low- and middle-income health systems stretch scarce clinicians by offering decision support, drafting records and answering health questions. Early evidence, such as a 2025 study in Nairobi clinics, shows fewer clinical errors. But models trained mostly on high-income data can recommend care that local systems cannot deliver, unless they are adapted to local guidelines, medicines and diseases.
A clinical officer in a rural district hospital types a question into a chatbot on her phone: an adult with fever, headache and a stiff neck, what should she do? The answer is fluent, confident and textbook-perfect. Arrange a CT scan before lumbar puncture. Send blood cultures. Start a third-generation cephalosporin plus vancomycin. Consider admission to intensive care.
Her hospital has no CT scanner. Blood cultures go to a laboratory four hours away and come back in a week, if at all. Vancomycin is not on the district formulary. There is no intensive care unit. The answer was correct. It was also useless, and in one respect dangerous: it may delay the one thing she can do, which is give the antibiotic she has, now.
This is the central tension of large language models in low- and middle-income health systems. They arrive with the promise of leapfrogging, giving every clinician a specialist in their pocket. But a model trained on the medicine of rich countries does not just transfer knowledge; it transfers assumptions about what is available. Whether LLMs leapfrog specialist shortages or quietly import someone else’s standard of care depends on who adapts them, and to what.
Why are LLMs attractive for low- and middle-income health systems?
The case is not hype. It rests on a real and worsening shortage. WHO projects a global shortfall of roughly 10 to 11 million health workers by 2030, concentrated in low- and lower-middle-income countries, and specialists are the scarcest of all. In many districts, the most senior clinician is a clinical officer or nurse, and a paediatrician or psychiatrist may be a day’s travel away. The loss of trained staff to richer countries, a theme we explored in our piece on why “brain circulation” is a comforting story told by destination countries, makes the gap harder to close with training alone.
LLMs are cheap per query, work in many languages, and can be delivered on a basic smartphone. They can help with tasks that eat clinical time: drafting referral letters, summarising histories, checking drug doses and explaining a guideline. In principle, they can put a second opinion in every consulting room.
And unlike many digital health products, they do not require building a new system first. That is exactly what makes them both promising and risky.
Does AI decision support actually reduce errors in African clinics?
The most cited real-world evidence comes from Kenya. In 2025, OpenAI and Penda Health, a primary care provider in Nairobi, published a study of AI Consult, a GPT-4o-based tool that ran in the background of consultations and flagged potential errors. Across 39,849 patient visits in 15 clinics, clinicians with the tool had a 16% relative reduction in diagnostic errors and a 13% relative reduction in treatment errors, as judged by independent physicians reviewing a sample of visits. The study had not been peer reviewed when released.
Two features of the Penda tool are easy to miss and essential to its results. It was adapted to Kenyan clinical guidelines, rather than relying on whatever the base model had learned. And it was designed to stay quiet unless something looked wrong. In other words, the local organisation shaped the model to its own medicine and its own workflow, not the other way round.
In January 2026 the Gates Foundation and OpenAI announced Horizon1000, a $50 million programme to bring AI tools to 1,000 primary care clinics and their communities in Africa by 2028, starting in Rwanda. It is an early sign that LLM deployment in low-income health systems is moving from pilot to programme, which makes the question of whose medicine the models encode more urgent, not less.
How well do LLMs understand African and other non-Western medical contexts?
The best systematic evidence so far is AfriMed-QA, a benchmark presented at the ACL 2025 conference, where it won a best social impact paper award. Researchers assembled about 15,000 medical questions, sourced from more than 60 medical schools across 16 African countries, covering 32 specialties. It includes expert multiple-choice questions, open-ended clinical questions and consumer health queries.
The results were telling. Every model tested scored lower on AfriMed-QA than on the US licensing-exam-style MedQA benchmark. The drop for GPT-4o was nearly 9 percentage points; for the smallest model tested, it was over 15. Specialised biomedical models did worse than general-purpose ones, and small models that could run on a phone struggled to reach a passing score. Interestingly, in human evaluations, consumers tended to prefer LLM answers to clinicians’ answers, which is not the same as those answers being right for the setting.
The gap is not mysterious. US exam questions assume US epidemiology, US drug availability and US diagnostic pathways. Questions written by African medical educators assume malaria in the differential, limited imaging, and national formularies. A model that aces one and slips on the other has learned a particular medicine, not medicine in general.
What does “importing someone else’s medicine” look like in practice?
It shows up in at least four ways.
Unavailable resources. As in the opening example, advice that assumes CT scanners, specialist referral, second-line drugs or intensive care can push clinicians toward care they cannot provide, or make them doubt the care they can.
Wrong disease priors. A model that has seen far more text about conditions common in high-income countries may rank them higher in a differential and underweight tropical infections, sickle cell complications or severe malnutrition.
Guidelines that do not match. National treatment guidelines differ, often for good reasons tied to resistance patterns, costs and drug registration. A model that recommends the US first-line antibiotic when the national guideline chooses something else creates a conflict the clinician must resolve on the spot.
Cost blindness. Advice that ignores what patients pay out of pocket can push families toward tests and medicines they cannot afford, and leave them poorer without leaving them healthier.
None of these is a failure of intelligence. They are failures of context. And context is precisely what a model trained on the internet’s medicine lacks for most of the world.
A worked example: adapting a model to a district formulary
This is a hypothetical example, built to illustrate what adaptation involves.
A ministry of health wants to deploy an LLM assistant for clinical officers in district hospitals. It starts with a general-purpose model and tests it on 300 real, anonymised cases from its own facilities, reviewed by local senior clinicians.
The first round is sobering. The model’s reasoning is often sound, but about a third of its recommendations involve a test or drug not available at district level, and it frequently cites international rather than national guidelines.
The ministry then does three things. It gives the model the national standard treatment guidelines and the essential medicines list to draw on for every answer. It tells the model which tests and drugs each level of facility has. And it asks the model to state, for every recommendation, what to do if the ideal option is unavailable.
On a second test set, the share of recommendations that local reviewers judge actionable at district level rises substantially. More important, the answers now start with what the clinician can do in the next hour. The model got no smarter. It got local.
The cost was not the model. It was the senior clinicians’ time to build the test set and review outputs, and a data team to keep the guidelines current. Those are the resources a ministry needs to budget for, and the ones vendors rarely mention.
The strongest counterargument: is “imported medicine” just a new excuse for doing nothing?
A fair critic would say that worrying about imported assumptions sets an impossible standard. Clinicians in low-income settings already rely on textbooks, apps and online resources written for high-income countries. They are skilled at filtering advice for local conditions. An LLM that is 90% right and occasionally recommends an unavailable test is still far better than no second opinion at all, which is the realistic alternative in many districts. Demanding bespoke local validation before any use could delay benefits for years while people die of treatable errors.
This has real force. The Penda evidence suggests benefit is achievable now, and waiting for perfect local models would be its own kind of harm.
But the argument cuts both ways. Penda’s gains came from a tool that was deliberately adapted to local guidelines and workflow, not from a raw chatbot. The evidence supports adaptation, not avoidance. The realistic choice is not between perfect local models and nothing. It is between models that are adapted with modest, deliberate effort and models that are dropped in as-is. The first path is affordable. The second quietly shifts the cost of mismatch onto clinicians and patients.
What this means for ministries, donors and developers
For ministries of health, the key is to own the context. Publish national treatment guidelines and essential medicines lists in machine-readable form. Require that any clinical LLM deployed in public facilities uses them and is tested on local cases, reviewed by local clinicians, before scale-up. WHO’s January 2024 guidance on large multi-modal models, which set out more than 40 recommendations for governments, developers and providers, gives ministries a framework to point to.
For donors and large programmes like Horizon1000, fund the unglamorous parts: local test sets, clinician reviewers, guideline maintenance and independent evaluation of patient outcomes, not just usage. Ensure local institutions hold the evaluation data and can repeat it when models change.
For developers, measure performance on local benchmarks such as AfriMed-QA as well as US exams, and report the gap. Build in the question “what if the ideal option is not available?” as a default, not an afterthought. As we argued in AI diagnostics: the evidence is strongest where the question is narrowest, narrow, well-specified tasks are where early deployments are most likely to earn trust.
Frequently asked questions
Can ChatGPT be used by doctors in developing countries?
Clinicians in many countries already use general-purpose chatbots informally. They can help with explanations, drafting and checking reasoning, but they are not validated clinical tools and may recommend tests or drugs unavailable locally. For clinical decisions, tools adapted to national guidelines and tested on local cases are safer. Clinicians should treat chatbot output as a second opinion to verify, not an instruction.
What is AfriMed-QA?
AfriMed-QA is a benchmark of about 15,000 medical questions sourced from more than 60 medical schools across 16 African countries, covering 32 specialties. Presented at ACL 2025, it found that all tested language models performed worse on it than on a US-style medical licensing benchmark, showing that high scores on Western exams do not guarantee performance in African clinical contexts.
Does WHO have guidance on large language models in health?
Yes. In January 2024 WHO published guidance on the ethics and governance of large multi-modal models in health, a category that includes large language models. It sets out more than 40 recommendations for governments, technology companies and health care providers, covering areas such as regulation, transparency, liability and assessment of risks before deployment.
How can LLMs be adapted to local health systems?
The most practical approaches give the model access to national treatment guidelines, essential medicines lists and information on what each facility can provide, then test it on real local cases reviewed by local clinicians. Designing prompts and interfaces to ask what to do when the ideal option is unavailable also helps. Adaptation depends less on retraining the model than on supplying context and checking outputs.
The takeaway
LLMs can be the specialist that most district hospitals will never have, but only if someone teaches them which hospital they are in. The leapfrog is real, and it is paid for in local guidelines, local test cases and local clinicians’ time. Skip that bill, and the model will practise someone else’s medicine, fluently, in your wards.
Have you seen an AI tool give advice that was right for a different health system? What happened next?
