There is a number that stops me every time I encounter it: according to the World Health Organisation, the world faces a projected shortfall of 10 million health workers by 2030, with the deficit concentrated overwhelmingly in low- and middle-income countries [1]. In Sub-Saharan Africa, there are regions where a single physician serves tens of thousands of patients. In rural South Asia, maternal mortality rates remain catastrophically high not because the knowledge to prevent those deaths does not exist, but because the human capacity to deliver care consistently does not.
Into this gap, artificial intelligence has arrived with considerable promise. AI triage systems — algorithms trained to assess symptom severity, prioritise patients, and flag urgent cases — are increasingly being piloted in low-resource settings as a partial answer to this workforce crisis. I want to engage with that promise seriously, because I think it is real. But I also want to be honest about something the technology press rarely says clearly: we are a long way from deploying these systems equitably. And until we confront why, the gap between promise and reality will widen, not narrow.
What AI Triage Can Do
Let us start with what is genuinely impressive. Modern AI triage systems built on large language models, decision tree ensembles, and symptom-checking neural networks have demonstrated meaningful clinical accuracy in controlled settings. Tools like Ada Health and Babylon Health have been evaluated against clinical benchmarks with results that, for common conditions, approach the diagnostic accuracy of a general practitioner [2]. In community health worker programmes — where frontline health workers with limited formal training conduct initial assessments — AI-assisted decision support has been shown to increase the accuracy of referral decisions and reduce dangerous undertriage [3].
The use case that excites me most is not replacing clinicians but extending them. A community health worker in a rural clinic, equipped with an AI-assisted triage tool on a basic smartphone, can systematically screen patients against a broader differential than memory and training alone would allow. The AI does not make the decision. It surfaces possibilities, flags red flags, and creates a structured record. The human makes the call. That division of labour is both clinically sound and ethically defensible — and in settings where the alternative is no screening at all, the marginal value is enormous.
"An algorithm trained on data that does not include you will not serve you well. It will serve the population it learned from — and call that universal."
The Problem Nobody Wants to Name
Here is where I need to say something uncomfortable. The datasets on which most clinical AI systems are trained are not representative of the populations in low-resource settings that these systems are now being asked to serve. They are drawn overwhelmingly from high-income, predominantly white patient populations in North American and European healthcare systems [4]. The result is a set of models that encode, with mathematical precision, the biases of the data they learned from.
This is not a hypothetical risk. It has been documented repeatedly. A landmark 2019 study published in Science found that a widely used clinical algorithm systematically underestimated the health needs of Black patients relative to white patients with identical clinical profiles — because the algorithm used historical healthcare spending as a proxy for health need, and Black patients had historically been given less care [5]. The bias was not in the intention. It was in the data. And the data reflected a healthcare system that had been inequitable for generations.
Extend this logic to AI triage in low-resource settings and the implications become serious. A model trained on the clinical presentations of patients in Boston or Berlin will have learned patterns that do not generalise to patients in Lagos or Dhaka — where disease burden, presentation timing, comorbidity profiles, and access to prior care are fundamentally different. Deploying that model as though it were universal is not neutral. It is a choice to embed existing inequity into algorithmic infrastructure.
Systemic Bias Is Not a Bug to Be Patched
I want to push back on a framing that I encounter frequently in AI ethics discussions: the idea that bias in clinical AI is a technical problem with a technical fix. Collect more diverse data, retrain the model, audit the outputs — and the problem is solved. This framing is dangerously incomplete.
Algorithmic bias in healthcare is not primarily a data problem. It is a reflection of structural inequity that predates machine learning by centuries. When a model underserves women, it is because medical research has historically underenrolled women in clinical trials [6]. When it underserves Black or brown patients, it is because those patients have been undertreated, underdiagnosed, and underrepresented in the clinical literature on which training data depends. When it fails in low-income settings, it is because the global health research apparatus has been oriented towards the diseases and presentations of wealthy populations.
More data from marginalised communities helps. But it is not sufficient on its own if the people collecting that data, labelling it, and making deployment decisions remain predominantly from the same demographics that produced the original bias. Equitable AI in healthcare requires equitable participation at every stage of the development process — not just in the dataset, but in the research teams, the clinical advisory boards, the ethics review processes, and the communities whose health is at stake.
Where bias enters the AI triage pipeline:
01 — Training data: sourced from high-income, predominantly white clinical populations — presentations, comorbidities, and disease burden differ significantly in low-resource settings
02 — Label quality: clinical ground truth labels inherit historical diagnostic biases — if physicians historically underdiagnosed conditions in certain groups, the model learns to do the same
03 — Proxy variables: using healthcare utilisation, spending, or prior diagnosis as proxies for health need encodes access inequity as clinical fact
04 — Deployment context: models validated in high-resource settings are applied in low-resource environments without revalidation against local epidemiology
What Responsible Deployment Actually Looks Like
None of this means AI triage should not be developed for low-resource settings. It means it should be developed differently — and that the communities it is intended to serve should be at the centre of that process from the beginning, not consulted at the end.
The most promising frameworks I have encountered treat AI triage not as a product to be exported from high-income to low-income settings, but as an infrastructure to be built collaboratively with local health systems. The WHO SMART Guidelines initiative [7] represents one of the most credible attempts at this — encoding clinical recommendations as computable, context-adaptable logic that can be implemented and validated locally, rather than a monolithic model trained elsewhere and deployed universally.
From a project management perspective — and my CAPM® training is relevant here — the failure mode of most AI health deployments in low-resource settings is not technical. It is structural. Scope is defined by the technology team, not the health system. Stakeholders from the target community are identified late in the process, if at all. Risk registers do not account for the social and political dimensions of deploying algorithmic decision-making in communities with legitimate reasons to distrust institutions. These are not AI problems. They are programme management problems, and they are solvable with the right frameworks applied from the start.
The First Step Is Acknowledgement
I want to end on the point I feel most strongly about, because it is the one most frequently skipped in technical discussions of this topic. Before better data, before fairer models, before responsible deployment frameworks — the first step is acknowledging that the problem exists and that it is structural.
The global AI health community has a tendency to frame equity as an optimisation target: something to be improved incrementally, measured in disparity metrics, and reported in evaluation tables. That framing is not wrong, but it is insufficient. It treats the underrepresentation of minority, low-income, and Global South populations in clinical AI as a technical gap to be closed rather than a symptom of a deeper problem — which is that the people building these systems have not, historically, been the people they are intended to serve.
Truly equitable AI triage in low-resource settings will not come from better algorithms alone. It will come from funding research led by scientists in the Global South. From clinical trials that enrol the populations whose lives depend on the results. From regulatory frameworks that require local validation before deployment, not after. From technology teams that reflect the demographic diversity of the patients their systems assess.
The technology is improving. The systems are becoming more capable. But capability is not equity. And until the field confronts the structural roots of its own bias — not just in the data, but in who holds the resources, the publishing power, and the deployment decisions — we will keep building faster, smarter tools that serve the already-served better than anyone else.
That is not good enough. And I think we know it.
Key Takeaways
Future Trends
The trajectory of AI triage in low-resource settings over the next decade will be shaped by these converging developments:
References