Seven Blind Spots in AI for Global Health
Attention, capital and evidence in AI for global health are concentrated in a narrow band of use cases: clinical decision support, image-based diagnostics and conversational tools.
Attention, capital and evidence in AI for global health are concentrated in a narrow band of use cases: clinical decision support, image-based diagnostics and conversational tools. These matter. But they are not where the gap between potential population-health benefit and current investment is largest.
My colleagues at EPIC Health Systems and I conducted evidence review to ask a specific question: in low- and middle-income countries (#LMICs), where is AI’s plausible contribution to health large, while investment, evidence or institutional capacity remains thin? We identified seven domains:
- The population data systems on which AI models depend
- Outcome-level evaluation of tools already being deployed
- Local languages, datasets and benchmarks
- Disease areas where burden and AI attention diverge
- Health-system functions behind the clinic: supply, financing, priority setting
- Regulatory and post-market surveillance capacity
- Second-order effects on labour markets, fiscal space and the information environment
Context: AI is arriving faster than the evidence
AI has moved quickly from pilots into routine health services in low- and middle-income countries. Large language models already support clinicians in primary care in Kenya. OpenAI and the Gates Foundation’s Horizon 1000 aims to reach 1,000 primary health care clinics by 2028, starting in Rwanda. Governments are also signing direct agreements with AI developers, such as Rwanda’s partnership with Anthropic covering health and the wider public sector. WHO has issued guidance on large multimodal models with more than 40 recommendations.
The field is developing unevenly. A Rethink Priorities landscape review found that only about a third of the frontier-AI organisations it identified are active in low- and middle-income countries, and that “frontier AI tools rarely align with strong evidence”. Its exploratory cost-effectiveness estimates varied by more than three orders of magnitude. When expected impact varies that much, where effort is directed matters as much as the technology itself. The seven domains below are where, on our reading, that choice is most often missed.
1. The data foundations AI depends on are eroding
Most AI and geospatial models in global health are calibrated against a small number of population data sources. When the US terminated funding for the Demographic and Health Surveys (DHS) Program in early 2025, surveys were set up or in the field in 23 countries; the programme had cost about $47 million a year. A Gates Foundation emergency grant of $39 million, with $4 million from Bloomberg Philanthropies, has kept core survey operations running, but the grant ends in 2028, the programme is narrower in scope.
The dependency is quantifiable. According to UN Population Division figures reported by PRB, DHS supplied 57% of the under-five mortality input data, 48% of adult mortality data and 24% of fertility data for least developed countries in World Population Prospects 2024. IHME’s child growth failure maps, for example, were built on 460 geo-referenced household surveys, drawn mainly from DHS, MICS and other national household surveys.
The alternatives are weak. Only 51% of children under five in sub-Saharan Africa have their births registered. During the COVID-19 pandemic, only four of 47 WHO African Region member states could report timely mortality data. Routine health information systems are not a substitute either. A four-country facility study found that concordance between register recounts and summary reports ranged from 9% to 40%. In Tanzania, DHIS2 data from 35 regional and tertiary hospitals recorded just over 8,500 neonatal deaths a year against about 48,000 expected nationally.
The risk is not only that estimates become less accurate. It is that model outputs remain precise-looking while their empirical anchors age, and continue to guide resource allocation. This domain is neglected because survey infrastructure sits outside the conventional definition of “AI investment”, even though every downstream model inherits its quality. There is also an opportunity: ICF reports early AI tools that automate DHS data processing, assist report writing and support natural-language data queries, which could lower the marginal cost of each survey round if validated.
2. Outcome-level evidence lags deployment
The evidence base for clinical AI remains thin and geographically concentrated. A scoping review of randomised trials of AI in clinical practice identified 86 RCTs, of which 31% were conducted in the United States and 30% in China; trials in Rwanda and Malawi each represented about 1% of the total. A 2026 review reports that fewer than 5% of FDA-cleared AI devices have peer-reviewed evaluation.
The most relevant LMIC case is also a cautionary one. An observational quality-improvement study, released as a preprint co-authored by OpenAI, of an LLM-based clinical copilot at Penda Health in Nairobi, covering 39,849 visits, reported 16% fewer diagnostic errors and 13% fewer treatment errors, with no statistically significant difference in patient-reported outcomes. The subsequent pragmatic cluster-randomised trial, with 9,691 patients across 16 facilities, found that LLM assistance “was safe but did not reduce treatment failure within 14 days”. A companion audit found harmful recommendations in 7.8% of encounters and noted “the asymmetric adoption of harmful versus beneficial outputs”.
A second signal comes from public-facing use. In a preregistered randomised study of 1,298 participants, LLMs tested alone identified relevant conditions in 94.9% of cases, but participants using the same models did so in fewer than 34.5% of cases, no better than controls. Benchmark performance does not transfer automatically to human-in-the-loop performance.
The gap is therefore between process measures (documentation, guideline adherence, error rates on audit) and patient and economic outcomes. Once a tool is embedded in routine care, the binding constraint is often less the model than what can be measured: missed visits, changed recording practices and thin follow-up data can make process gains look like health gains. EVAH’s $60 million commitment to locally led evaluation is a welcome correction.
3. Languages, datasets and benchmarks still reflect high-income contexts
Language is a clinical variable: history-taking, counselling and adherence support all depend on it. Yet in the latest Common Crawl snapshot, English accounts for 41.9% of pages, while Swahili accounts for 0.011%, Amharic 0.005% and Hausa 0.003%. By our summation, all sub-Saharan African languages that the crawler’s detector can identify together make up roughly 0.05%. The detector recognises 160 languages; Africa has over 2,000. Evaluations confirm the consequence: IrokoBench reports a “significant performance gap” between high-resource languages and 17 African languages, and LLM health responses differ markedly even across major languages such as Spanish, Chinese and Hindi.
Benchmarks are improving but remain partial. AfriMed-QA, built from 15,000 questions drawn from over 60 medical schools in 16 countries, found that model performance lags US licensing-exam benchmarks and that smaller, edge-deployable models struggle to reach a passing score. It is, however, English-only. OpenAI’s HealthBench involved 262 physicians who have practised in 60 countries and includes a global-health theme, but its conversations are partly synthetic and its grading is model-based.
The underlying training data share the same skew. In a review of 7,314 clinical AI papers, the US and China supplied more than half of the datasets, and radiology was the most represented specialty. Across 21 open-access skin-cancer image datasets, Fitzpatrick skin type was recorded for only 2.1% of images. This domain is neglected because local-language corpora, LMIC-specific benchmarks and clinical datasets are public goods: costly to build, easy to free-ride on, and without a commercial owner. Efforts such as the Lacuna Fund (39 published datasets to date) show the model works, but not at the scale the problem requires.
4. Disease burden and AI attention diverge
Where AI products concentrate largely reflects where commercial markets and available data pull. By our count of the FDA’s public list, radiology was the lead review panel for about 76% of AI-enabled device authorisations between January 2024 and June 2026. That is a US market signal, but it shapes which products are available to adopt elsewhere. Research portfolios are more mixed: among the 97 LMIC-focused NIH AI grants, 72% targeted diagnostics and treatment (versus 67% of grants focused on high-income countries), although non-communicable diseases were the most common disease area (39%).
The more useful question is therefore not whether a disease is studied, but whether AI is addressing the functions that determine outcomes where burden is highest:
- Non-communicable diseases. NCDs caused at least 43 million deaths in 2021, and 82% of the 18 million premature NCD deaths occurred in LMICs. Of 1.4 billion people with hypertension, just over one in five have it under control. The constraint is longitudinal: continuity, titration and adherence, which current AI tools rarely address.
- Mental health. Over a billion people live with a mental health condition. Government mental health spending ranges from $65 per person in high-income countries to $0.04 in low-income countries. We found little rigorous evaluation of AI-supported psychological interventions in LMIC settings.
- Maternal and newborn health. There were 2.3 million neonatal deaths in 2024 and 1.9 million stillbirths in 2023.
The maternal case also shows what rigorous LMIC-partnered AI work can look like. A Gates-funded model estimating gestational age from “blind sweep” ultrasound achieved a mean absolute error of 3.9 days versus 4.7 days for standard biometry in a US–Zambia test set, and performance held when untrained providers in Zambia collected sweeps on low-cost devices (4.9 versus 5.4 days). A later prospective study confirmed equivalent performance by novice users. In many LMICs, nearly half of pregnant women receive no scan. Yet the same review found only 14 eligible studies, two of them reviews. One caveat applies to this whole domain: we found no source that quantifies AI investment by disease area in LMICs. That absence is itself a gap.
5. Health-system functions behind the clinic
Most AI-for-health effort targets the clinical encounter. Yet substantial efficiency losses in LMIC health systems plausibly occur upstream: in commodity supply, workforce deployment, purchasing and priority setting. The Rethink Priorities landscape review’s impression was that more than two-thirds of organisations working on clinical workflow improvement or medical-record analysis are active only in high-income countries, and that commercial demand for LMIC-focused service-delivery tools is limited.
The workforce arithmetic makes this consequential. WHO projects a shortfall of 11 million health workers by 2030, mostly in low- and lower-middle-income countries. Under that constraint, tools that reduce documentation and reporting burden on frontline and community health workers can release more clinical time than tools that marginally improve diagnostic accuracy, but they attract less attention because their gains are diffuse and harder to attribute. The same review cautions that efficiency tools often fail to convert time savings into health gains unless clinician time is the binding constraint; the workforce shortfall makes that condition common.
Priority setting is the most under-examined function. Health technology assessment agencies are beginning to define how AI may be used in evidence generation; NICE published a position statement in 2024. To our knowledge, comparable guidance for LMIC priority-setting institutions does not yet exist. From a health-economics standpoint, this is a significant omission. Better allocative decisions across a whole benefit package, through faster evidence synthesis, adapted cost-effectiveness models and claims analytics, could plausibly yield larger health gains per dollar than improving technical efficiency within a single clinic. That proposition is testable and, as far as we can find, largely untested.
6. Regulatory and post-market surveillance capacity
AI tools are entering LMIC health systems faster than the institutions that would assure them. On WHO’s Global Benchmarking Tool list as of August 2026, nine African national regulatory authorities operate at maturity level 3 for medicines, and none at level 4. Globally, only one authority, Singapore’s, is listed with medical-device scope. On our reading, no African regulator has yet been benchmarked for the product category into which most clinical AI falls.
Continental structures are young. The African Medicines Agency became operational in October 2025; by July 2026, 33 of 55 AU member states had ratified its treaty. It focuses on vaccines and complex products. WHO’s 2023 document on regulating AI for health states explicitly that it is “not intended as guidance”. The AU’s Continental AI Strategy, endorsed in July 2024, sets direction but not product-level assurance.
Post-market performance is where the gap bites. Computer-aided detection for tuberculosis is among the best-evidenced AI tools in global health and has a WHO recommendation. Even so, a Ghanaian study found that a locally calibrated threshold dominated current practice and was cheaper than the vendor default, at the cost of slightly lower effectiveness. Threshold choice is therefore a local economic decision, not a technical setting. Chest X-rays from Southern African ex-miners formed a distinct cluster separate from international images. If threshold calibration and drift monitoring matter for a mature, single-task tool, they matter more for general-purpose LLMs. Few LMIC health systems currently have the mandate, staff or funding to do either.
7. Second-order effects: labour, fiscal space and the information environment
The largest health effects of AI in LMICs may not run through health care at all. Income, employment and public revenue are among the strongest determinants of population health, and AI will affect each. The IMF estimates that almost 40% of global employment is exposed to AI, compared with 26% in low-income countries. The ILO’s refined index finds that one in four workers globally is in an occupation with some generative-AI exposure, with clerical work the most exposed. Lower average exposure in poorer economies is not reassuring if it is concentrated in the formal, urban, service-export jobs that underpin tax revenue and middle-class growth.
Fiscal space is already tight. UNCTAD reports that 3.7 billion people live in developing countries where interest on public debt exceeds public spending on health or education, and that the number of such countries rose from 43 in 2021 to 51 in 2023. An AI-related shock to formal employment or the tax base would fall on health budgets that have little room left to absorb it, at a moment when external assistance is also falling.
The information environment is the third channel. As general-purpose chatbots become a first point of health information, the evidence that users do not reliably extract correct advice from them (Section 2) becomes a population-level concern, particularly in languages where model performance is weakest (Section 3). These pathways are plausible and consequential. They are also largely unmeasured, and we found little health funding directed at measuring them.
Implications
The common thread across these seven domains is that the neglected work is infrastructural, public-good in character and slow to show attributable results. That is precisely the profile that commercial markets and short grant cycles under-supply. Six practical implications follow.
- Treat population data systems as AI infrastructure. Funders of AI deployment should co-finance survey continuity and civil registration, and test whether AI can lower the cost of each survey round without loss of validity.
- Make outcome evaluation a condition of scale, not an afterthought. Large deployments should publish a protocol, prespecify patient and cost outcomes, and budget for independent evaluation from the outset.
- Fund local-language datasets and LMIC benchmarks as public goods, governed by institutions in the countries they describe, and track frontier-model performance on them over time.
- Measure, then rebalance. Funders should track AI investment by disease area and health-system function, and test whether hypertension control, mental health and maternal-newborn care, alongside supply chains, workforce support and priority setting, receive attention commensurate with their burden.
- Invest in regulatory science and post-market monitoring, working through the African Medicines Agency and maturity-level-3 regulators, with local calibration treated as a routine requirement rather than a research project.
- Bring second-order effects into health planning. Ministries of finance and health should model AI exposure in labour markets and revenue when setting medium-term health budgets and safety-net policy.
Conclusion
The question is no longer whether AI will reach health systems in LMICs; it already has. The more consequential question is whether the surrounding infrastructure will be able to tell us if it is working, for whom and at what cost. On current evidence, the most valuable marginal dollar may lie less in the next clinical application than in the data, evaluation, regulatory and analytical capacity that determines whether any application delivers health gains.
The views expressed are the author’s own.
Systems and Sovereignty by email
New issues, sent when they are published.













