In 2020, a Google DeepMind paper published in Nature demonstrated that an AI system called AlphaFold had effectively solved protein folding — one of biology’s hardest long-standing problems. In 2023, AI systems began matching or exceeding radiologist performance in identifying breast cancer, lung cancer, and diabetic retinopathy from medical images. In 2024, AI tools began assisting drug discovery in ways that reduced the timeline for identifying promising compounds from years to months.
These are not marginal improvements. They represent a qualitative shift in what medicine can do.
The promise of AI in healthcare is, at its best, transformative. AI does not get tired. It does not have cognitive biases that cause it to anchor on the first diagnosis that comes to mind. It can process patient histories, genetic data, imaging results, and population-level evidence simultaneously in ways that exceed any individual clinician’s cognitive capacity. At scale, AI-assisted diagnostics, personalized treatment selection, and predictive risk identification could avert millions of deaths per year from diseases that are currently caught too late, misdiagnosed, or undertreated.
And yet the same capabilities that make AI powerful in healthcare also make it uniquely threatening to the privacy and autonomy of patients. Health data is among the most sensitive data that exists — it reveals not only current conditions but vulnerabilities, genetic predispositions, mental health histories, reproductive decisions, and risk profiles for conditions not yet developed. The infrastructure required to train, validate, and operate AI healthcare systems at scale is also the infrastructure required to surveil, discriminate against, and control people in ways that healthcare ethics has spent decades trying to prevent.
The question is not whether AI will transform healthcare. It will. The question is whether the transformation is governed in a way that delivers benefits to patients — all patients, not just those in well-resourced settings — while protecting the rights and interests that medicine is supposed to serve.
What AI Is Already Doing Well
The clinical evidence base for AI in healthcare has matured rapidly. Several applications have demonstrated genuine performance advantages over standard care.
Radiology and imaging: AI systems trained on large datasets of labeled medical images can identify pathologies — tumors, fractures, retinal disease, skin lesions — with accuracy that matches or exceeds specialist radiologists in controlled conditions. Google’s ARDED system for diabetic retinopathy detection has been validated across multiple health systems in the UK, US, India, and Thailand, and reduces the burden on scarce ophthalmologists while identifying disease at stages when treatment is most effective. The FDA has approved dozens of AI-assisted imaging tools; European regulatory approvals are similarly accumulating.
Sepsis prediction: Sepsis — life-threatening infection that kills approximately 11 million people annually — is one of medicine’s most time-sensitive emergencies. AI systems trained on electronic health records can identify patients at high risk of sepsis hours before clinical signs become obvious to clinicians, enabling earlier intervention. Hopkins Hospital’s AI sepsis prediction tool has been associated with significant mortality reductions in multiple studies.
Drug discovery: AI modeling of protein interactions, molecular docking, and compound activity has shortened the early-phase drug discovery pipeline in ways that multiple pharmaceutical companies are now operationalizing. The impact on rare disease, in which the economics of traditional drug development made many conditions commercially non-viable to target, is potentially transformative.
Clinical documentation: One of the most mundane but significant AI applications is clinical documentation support — AI tools that listen to consultations, generate clinical notes, and reduce the time clinicians spend on administrative documentation. Clinician burnout driven by documentation burden is one of the largest contributors to healthcare workforce attrition in many countries; AI documentation tools address a real and severe problem.
Where AI Fails — and the Bias Problem
The same datasets that enable AI performance also embed the biases of historical medical practice.
AI systems trained primarily on data from high-income country health systems, which historically underrepresented women, non-white patients, and elderly populations in clinical research, can perform substantially worse on the underrepresented groups. An AI skin cancer detection system trained predominantly on images of light skin tones will perform worse on darker skin tones — a well-documented failure in multiple commercial systems. An AI risk prediction system trained on US health records will not generalize reliably to health systems in South Asia or sub-Saharan Africa.
More subtly, AI systems can encode existing disparities in care. If a risk score for hospital readmission is trained on historical data in which Black patients were systematically undertreated — and were therefore less likely to be rehospitalized purely because they received fewer resources — an AI predicting readmission risk may learn to identify Black patients as lower-risk, reinforcing the undertreatment rather than correcting it. This type of algorithmic bias in healthcare is documented, consequential, and not easily detectable without deliberate auditing.
AI performance can also be brittle: a system trained on data from one hospital’s imaging equipment may not perform as well on data from a different manufacturer’s scanner, or from a different patient population. The validation conditions under which AI tools are approved may not reflect the deployment conditions in which they are used. The FDA and other regulators are developing “real-world performance” monitoring requirements, but these are new and not yet consistently implemented.
The Data Problem
Training AI systems capable of medical-grade performance requires vast quantities of high-quality labeled patient data — precisely the data that is most sensitive and most protected by medical privacy law.
The major AI companies in healthcare have access to this data through partnerships with major hospital systems. Google has partnerships with Mayo Clinic and dozens of other health systems. Microsoft, through its acquisition of Nuance (a clinical documentation AI company), has access to data from thousands of healthcare facilities. Amazon Web Services provides cloud infrastructure for a significant proportion of US health data. Apple Health data, drawn from hundreds of millions of iPhones, provides population-level health signals.
The concentration of health data in technology companies creates a structural power imbalance that patients, clinicians, and health systems are only beginning to grapple with. The terms under which technology companies access health data — what they can use it for, what protections apply, what commercial applications they can build — are negotiated in conditions of significant information asymmetry and limited patient agency.
In the United States, HIPAA provides baseline health data protections but was written before AI, was designed around individual clinician-patient relationships rather than population-scale data flows, and has significant gaps when it comes to data flowing to technology intermediaries. Europe’s GDPR provides stronger protections and is being supplemented by the AI Act’s specific requirements for high-risk AI systems in healthcare. The regulatory frameworks are not yet commensurate with the scale of data flows.
Patient consent for health data use in AI training is another underresolved question. Patients typically consent to their data being used for their care. Whether that consent extends to use in AI training programs — which may benefit other patients but not directly the patient whose data is used — is legally ambiguous and ethically contested.
Equity: Who Benefits from Medical AI?
Perhaps the deepest question about AI in healthcare is distributional: who captures the benefits?
AI systems that improve diagnostic accuracy in sophisticated referral hospitals in Boston or London have clear benefits for the patients those hospitals serve. Whether equivalent benefits flow to patients in rural clinics in Mississippi, or in district hospitals in Kenya or Bangladesh, depends on deployment decisions, infrastructure investment, and pricing that markets do not automatically optimize.
Health AI tools priced for high-income country healthcare systems are frequently unaffordable for health systems in low- and middle-income countries. The concentration of the best AI development capacity in high-income country companies and research institutions tends to produce systems trained on high-income country data that perform best on high-income country populations. The global burden of disease is disproportionately borne by low-income populations; the benefits of medical AI are disproportionately flowing to high-income populations.
There are exceptions and genuine counter-examples. Several AI diagnostic tools — including AI screening for tuberculosis, AI-assisted malaria diagnosis, and AI grading of diabetic retinopathy — have been specifically designed for deployment in low-resource settings and have demonstrated effectiveness in high-burden countries. These examples prove the possibility. They do not represent the default.
The equity dimension of AI in healthcare is not merely an ethical concern for those with the luxury of additional priorities. It is a public health priority: if the global burden of disease remains concentrated in low-income settings because advanced diagnostics and treatment selection tools are inaccessible to the health systems serving those populations, AI becomes a technology that widens global health disparities rather than reducing them.
Also explore:
Digital Privacy: The Complete Guide
Antibiotic Resistance: Humanity’s Next Silent Pandemic Is Already Here (published July 10)
The Digital Divide: Who Benefits from AI and Who Gets Left Behind (coming September 3)
Sources & Further Reading
- GDPR.eu — the General Data Protection Regulation explained clause by clause.
- Privacy International — investigations into state and corporate data practices.
- Stanford Encyclopedia of Philosophy — Privacy — the conceptual groundwork behind privacy claims.
- EFF — Surveillance Self-Defense — practical guidance on resisting tracking.
Related Articles
- The Erosion of Privacy in AI Government Surveillance
- Navigating the AI Consciousness Ethical Dilemma
- AI Existential Risk: The Hidden Dangers Ahead
Weekly Geopolitical Briefing
Go Beyond the Headlines
Analysis of the forces shaping the world - delivered every week. No noise, no bias, just depth.
Subscribe Free ?No spam. Unsubscribe anytime.
About the Author
António Monteiro
Engineer by profession, geopolitical analyst by conviction. I believe responsibility for the planet's future doesn't belong only to governments and institutions - it belongs to all of us. Knowledge about geopolitics, international conflicts, and the forces shaping the world is the most powerful tool for becoming more conscious, informed citizens. You don't need to be a diplomat to understand what's at stake - you just need to want to go beyond the headlines. At Outside The Case, I analyze conflicts, power dynamics, and global trends with rigor and accessible language, so you can understand what's really happening in the world.
Read more about the author →