In Malawi, one radiologist serves 8.8 million people. She has no AI tools, no algorithmic support, just one pair of hands and an overwhelming caseload. Across the world, hospitals in wealthy nations are already deploying algorithms to read those same scans, often without patients ever knowing. In 2015, the US Food and Drug Administration cleared just two AI-powered medical devices. By 2024, that number had surged past 250, a hundredfold increase in a decade, with a promise that faster diagnoses and fewer errors would follow.
But for whom was this revolution actually carried out?
Over 1,200 AI medical devices have been authorized by the FDA since 2015. Nearly all were developed in, and designed for, wealthy nations.
Behind each of those 1,200 authorizations is a real clinical deployment, with an algorithm quietly reading a scan in a Chicago hospital, flagging a cardiac anomaly in London, and screening a retina in Seoul. The growth is extraordinary. But the geography is not. Every single one of those deployments happened in a wealthy country. The chart tells a story of progress. What it cannot show is how narrow the focus has remained, or who was written out of it entirely.
More than three quarters of all AI medical devices do one thing: read medical images. Cardiology gets a small share, while pathology, ophthalmology, and neurology are barely represented.
This is not a broad healthcare revolution. It is a gold rush concentrated in a single corridor, always dug where digital infrastructure already runs deep. Hospital-grade scanners, high-bandwidth networks, and vast imaging datasets: these prerequisites exist almost exclusively in wealthy countries. The specialty concentration does not just reflect where the technology has advanced fastest. It reflects who was designed into the story from the start.
Who is digging that corridor? The answer is a remarkably small group.
A handful of corporations from the world’s wealthiest nations own this technology. Every single top manufacturer is headquartered in a high-income economy. Not one in Sub-Saharan Africa. Not one in South Asia.
The countries with the greatest health needs have zero seats at the table where healthcare AI is built. When a few American companies dominate, they build for American hospitals with American patient data. Diseases that primarily strike tropical countries, or conditions common in low-resource settings, simply aren’t profitable enough to build AI for.
The result is a world where the distribution of AI tools mirrors the distribution of wealth, not need.
This market logic has a name. In 1971, British physician Julian Tudor Hart described the Inverse Care Law: the availability of good medical care tends to vary inversely with the need for it. Half a century later, AI is writing a new chapter of the same story.
Nations with the greatest need, those with the highest mortality, fewest doctors, and lowest spending, have the least readiness to use AI. Sub-Saharan Africa’s entire cloud computing infrastructure is smaller than Switzerland’s.
Rather than bridging the gap, AI appears to be widening it. The data makes the cost concrete: Malawi registers an AI readiness score of just 0.10 and health expenditure of $27 per person per year, while serving one radiologist for every 8.8 million people. Without reliable internet, without electricity in rural clinics, without trained staff to act on AI outputs, these tools remain science fiction.
Yet even in wealthy countries where deployment is possible, a more unsettling question remains: how good is the evidence that these devices actually work?
The evidence base for these devices is alarmingly thin. Most cleared devices have never undergone a proper clinical trial, which remains the accepted gold standard of medical testing. We are running a live experiment on patients, with no control group and no informed consent about the risks.
More than 690 AI medical devices have been cleared through a single country’s regulatory process, yet real-world clinical trials remain the exception rather than the rule. Imagine a drug being prescribed to millions without ever being tested in a proper trial. That’s essentially what’s happening with AI medical devices right now.
AI may well transform medicine. But right now, these tools are being sold to the richest hospitals, tested on the narrowest populations, and cleared with the thinnest evidence. Readers, patients, and policymakers should be asking: who is checking whether these algorithms actually work, and for whom?
The answer, right now, is almost no one. The radiologist in Malawi is still waiting for tools that could save lives. Patients in wealthy hospitals are already using tools that have barely been tested. Both failures share the same cause: a system that prioritised speed to market over rigour, and profit over need.
Requiring diverse validation data for regulatory clearance, funding AI tools for neglected diseases, and demanding clinical trial evidence before deployment are not radical demands. They are simply the standards we already apply to every other medical technology. It is time to apply them to AI.
Food and Drug Administration. (2025). Artificial intelligence and machine learning (AI/ML)-enabled medical devices. U.S. Food and Drug Administration. https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-aiml-enabled-medical-devices
International Monetary Fund. (2023). AI Preparedness Index (AIPI). IMF DataMapper. https://www.imf.org/external/datamapper/datasets/AIPI
Tariq, A., Alfattal, S., Gichoya, J. W., & Banerjee, I. (2025). Current reporting on evidence used for US Food and Drug Administration regulatory clearance of artificial intelligence/machine learning-enabled medical devices. JAMA, 333(4), 342–344. https://doi.org/10.1001/jama.2024.23832
World Bank. (2024). Current health expenditure per capita (current US$) [Data set]. World Bank Open Data. https://data.worldbank.org/indicator/SH.XPD.CHEX.PC.CD
World Health Organization. (2024). Global Health Observatory data repository. World Health Organization. https://www.who.int/data/gho
Maru, S., Matthias, M. D., Kuwatsuru, R., & Simpson, R. J., Jr. (2024). Studies of artificial intelligence/machine learning registered on ClinicalTrials.gov: Cross-sectional study with temporal trends, 2010–2023. Journal of Medical Internet Research, 26, e57750. https://doi.org/10.2196/57750
Vota, W. (2026, May 28). Compute reality of artificial intelligence in global health LMICs. ICTworks. https://www.ictworks.org/compute-reality-of-ai-in-global-health-lmics/
All code, data analysis, and visualisation design in this project were completed independently by the author. No generative AI tools were used in this assignment.
Data was sourced from the US FDA, IMF, and World Bank open data repositories. Thanks to The Conversation Australia for the publication style guide.