Redistributed

A clinic in rural Tamil Nadu has no radiologist. The nearest one is four hours away. A patient arrives with a cough that won't quit and a chest X-ray taken on aging equipment. An AI system reads the scan in seconds, flags a suspicious mass, and recommends follow-up. In a well-funded Boston hospital, the same AI system runs alongside a team of specialists who barely need it. Both sites use the same algorithm. The outcomes couldn't be more different.

This is the central tension of AI in healthcare. The technology has genuine diagnostic power. It can detect patterns in medical images that human eyes miss, identify drug candidates in months instead of years, and reduce the paperwork drowning clinicians. But the question isn't whether AI can improve healthcare. It's who gets the improvement — and at what cost to everyone else. This is a redistribution question, and healthcare makes it concrete in ways that abstract debates about AI and Consumer Surplus cannot.

Where It Actually Works

AI in healthcare isn't speculative. Several applications have moved past proof-of-concept into clinical use, with measurable results.

Medical imaging is the strongest case. AI systems can now detect diabetic retinopathy from retinal scans with accuracy matching or exceeding ophthalmologists. Google's work on retinal imaging, validated across multiple clinical settings, demonstrated that an algorithm trained on roughly 130,000 images could screen for a condition that causes preventable blindness in millions of diabetic patients worldwide. The FDA has cleared over 500 AI-enabled medical devices, the majority in radiology — reading mammograms, CT scans, and pathology slides.

Meanwhile, DeepMind's AlphaFold cracked a problem that had stumped structural biologists for fifty years: predicting protein structures from amino acid sequences. The implications cascade through drug design, disease understanding, and biological research. AlphaFold has predicted structures for nearly every known protein — over 200 million of them — and made the database freely available. This is a genuine scientific breakthrough, and its open availability is worth noting because it's the exception, not the rule, in how AI's medical value gets distributed. It demonstrates what becomes possible when AI capabilities are treated as public infrastructure rather than proprietary advantage.

Drug discovery is accelerating. The traditional pipeline from target identification to approved drug takes 10-15 years and costs an average of $2.6 billion. AI companies like Recursion Pharmaceuticals, Insilico Medicine, and Isomorphic Labs (a DeepMind spinoff) are compressing the early stages — identifying candidate molecules, predicting their behavior, and filtering out failures before expensive clinical trials. Insilico brought an AI-discovered drug for idiopathic pulmonary fibrosis to Phase II clinical trials in roughly 30 months from target identification, a process that traditionally takes four to six years.

Clinical decision support systems now flag potential drug interactions, suggest differential diagnoses, and identify patients at risk of deterioration. Sepsis prediction algorithms, for example, can alert clinicians hours before a patient's vital signs visibly crash. These aren't replacing clinical judgment — they're adding a statistical layer underneath it.

Administrative AI may be the least glamorous application and the most immediately impactful for clinicians. Tools like Nuance DAX and Abridge listen to patient encounters and generate clinical notes automatically. Physicians spend roughly two hours on documentation for every hour of patient care. If AI halves that burden, the result isn't just efficiency — it's clinicians who can actually look at their patients instead of their screens.

Where It's Overpromised

For every working application, there's a marketing pitch that outpaces reality. This gap between capability and claim is a recurring pattern across AI (benchmarks rarely capture real-world reliability), but in healthcare the stakes are uniquely high.

The "AI doctor" — a chatbot that replaces clinical judgment — doesn't exist in any meaningful sense. Large language models can pass medical licensing exams. They can also hallucinate drug dosages, invent contraindications, and deliver a confident diagnosis based on pattern-matching that no clinician would accept without examination. As Meredith Broussard documents in More Than a Glitch, the gap between performing well on structured tests and functioning reliably in the messy reality of clinical practice remains vast.

Fully automated surgery is similarly distant. AI-assisted surgical systems like Intuitive's da Vinci robot are sophisticated tools controlled by surgeons, not autonomous operators. The hand, the judgment, and the liability remain human. Marketing materials that blur this distinction serve shareholders, not patients.

And the perennial promise of "personalized medicine" — genomic data combined with AI to tailor treatment to your specific biology — remains largely unrealized for most patients. It works in narrow oncology applications, where tumor genomics can guide drug selection. For the broad population, it's still more promise than practice.

The Redistribution Question

Here's where the diagnostic revolution meets redistribution. Who captures the value that AI creates in healthcare?

The hospital divide. AI diagnostic tools require capital — for hardware, software licenses, integration with electronic health records, staff training, and ongoing maintenance. The hospitals that can afford this are already the best-resourced ones. The Cleveland Clinic, Mayo, and Massachusetts General adopt AI tools that make their good care better. Meanwhile, safety-net hospitals serving low-income populations — already stretched thin — fall further behind. AI doesn't create the gap, but it widens it.

Private vs. public systems. In private healthcare markets like the United States, AI capabilities become premium features. "AI-enhanced diagnostics" join the menu of services that better insurance buys you. In public systems like the UK's NHS, adoption depends on government procurement budgets and political will. The NHS has invested in AI diagnostics — its AI lab launched in 2019 — but funding constraints mean deployment is uneven and slow. The result: AI accelerates the two-tier dynamic that already defines healthcare access in most countries.

Drug discovery economics. AI may cut the cost of identifying drug candidates, but nothing in the current pharmaceutical business model ensures those savings reach patients. If AI reduces early-stage drug development costs by 30%, pharmaceutical companies can either lower drug prices or keep prices high and pocket the margin. Every incentive in the system points toward the latter. Insilico's AI-discovered drug, if approved, will be priced by the same market logic that gave us $84,000 hepatitis C treatments and insulin costs that have tripled in two decades. The tool changes; the power structure doesn't — unless pricing and competition policy change with it (see AI and Antitrust for how market structure shapes these outcomes).

The global access gap. AI diagnostics could be transformative in settings where specialists simply don't exist. Sub-Saharan Africa has roughly one radiologist per million people, compared to roughly 100 per million in the United States. An AI system that reads chest X-rays could, in principle, bring specialist-level screening to clinics that have never had it.

But "in principle" is doing heavy lifting. These systems need reliable internet, compatible imaging equipment, electrical infrastructure, and local technical support. They need to be trained on populations they'll serve — and most aren't. As Abeba Birhane and collaborators have documented, AI systems built on data from wealthy-country populations often fail when deployed in different demographic contexts. A retinal scan algorithm trained on images from American and European patients may miss pathology that presents differently in South Asian or African populations.

The Bias Problem

The data problem in medical AI isn't theoretical. It's documented and consequential. It connects to the broader challenge of bias in AI systems — but in healthcare, the consequences are measured in lives.

Dermatology AI provides the starkest example. Multiple studies have shown that skin cancer detection algorithms trained predominantly on images of light skin perform significantly worse on dark skin tones. A 2021 study in The Lancet Digital Health found that only 16% of publicly available dermatology datasets included any labeling for skin tone, and those that did were overwhelmingly skewed toward lighter skin. An AI system that catches melanoma early for white patients but misses it for Black patients doesn't just fail — it actively widens a health disparity that already kills.

Clinical decision tools inherit biases from the medical records they're trained on. An algorithm widely used in US hospitals to allocate healthcare resources was found in 2019 to systematically underestimate the illness severity of Black patients. The reason: it used healthcare spending as a proxy for health needs. Because Black patients historically received less care (due to access barriers, insurance gaps, and discrimination), the algorithm concluded they were healthier. The bias wasn't in the math. It was in the history the math absorbed. As Ruha Benjamin argues in Race After Technology, these systems don't just reflect existing inequality — they launder it, giving it the appearance of computational objectivity.

Ziad Obermeyer, the researcher who identified that algorithm's bias, estimated it affected the care of millions of patients. The hospital system eventually adjusted the tool. But the episode reveals a structural problem: medical AI systems are being deployed at scale before their differential performance across populations is adequately tested.

What Equitable AI Healthcare Would Look Like

The technology itself is agnostic. It can widen or narrow disparities depending on how it's governed. What would the narrowing path look like?

Public investment in AI diagnostics for underserved settings. If AI chest X-ray readers can work in rural clinics, fund them as public health infrastructure — the way vaccines are funded, not the way MRI machines are sold. The WHO has begun developing frameworks for evaluating and deploying AI health tools in low-resource settings, but frameworks without funding are wishes.

Data diversity requirements. Before an AI diagnostic tool receives regulatory clearance, require evidence that it performs adequately across demographic groups. The FDA has moved in this direction but hasn't mandated it. Without mandates, companies optimize for the populations that generate revenue, not the populations that need care.

Open-source medical AI. AlphaFold's open database is the model. When publicly funded research produces AI tools with medical applications, those tools should be publicly available. The argument for proprietary medical AI is weaker than for proprietary pharmaceuticals — the marginal cost of distributing a trained model is essentially zero. Organizations like the Open Source Imaging Initiative, Hugging Face's medical model repositories, and partnerships between academic medical centers are exploring this space. The growing small language model ecosystem makes local deployment increasingly feasible for clinics that cannot rely on cloud APIs. These remain small against the commercial tide, but they represent a viable alternative path.

Regulatory standards that account for equity. Medical device regulation currently focuses on average performance: does the device work? The question should be: does it work for everyone? And if not, who does it fail? The EU's AI Act includes provisions for high-risk AI systems in healthcare, including requirements for testing across populations. Whether these translate from regulation to practice will determine their value.

The Fork in the Road

AI in healthcare presents a genuine fork. Down one path, AI becomes another layer of advantage for well-resourced health systems — better diagnostics for those who already have good diagnostics, faster drug development that still produces expensive drugs, clinical tools trained on and optimized for wealthy-country populations. This is the default path. It requires no policy intervention. It just requires inaction.

Down the other path, AI becomes a tool for narrowing health disparities — bringing specialist-level diagnostics to places without specialists, catching diseases earlier in populations that currently present late, freeing clinicians from paperwork so they can spend time with patients. This path requires active choices: public funding, data standards, open-source commitments, regulatory mandates, and the political will to treat AI healthcare tools as infrastructure rather than luxury goods.

The technology doesn't choose the path. The institutions deploying it do. And right now, the institutions deploying AI in healthcare are overwhelmingly driven by the same incentives that created the disparities AI could, in theory, help close.

The diagnostic revolution is real. The question is whether it's a revolution for everyone, or just another upgrade for those who can already afford the best care.

Sources

Birhane, Abeba. "The Impossibility of Automating Ambiguity." Artificial Life, 2021. MIT Press
Benjamin, Ruha. Race After Technology: Abolitionist Tools for the New Jim Code. Polity, 2019. Publisher
Broussard, Meredith. More Than a Glitch: Confronting Race, Gender, and Ability Bias in Tech. MIT Press, 2023. Publisher
Daneshjou, Roxana, et al. "Disparities in Dermatology AI Performance on a Diverse, Curated Clinical Image Set." Science Advances, 2022. Science
Obermeyer, Ziad, et al. "Dissecting Racial Bias in an Algorithm Used to Manage the Health of Populations." Science, 2019. Science
Topol, Eric. Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again. Basic Books, 2019. Publisher

General

AI in Healthcare: Who Benefits from the Diagnostic Revolution