When More Medicine Does Not Mean More Health, Governance Becomes the Treatment

Jul 16, 2026

10 min read

5 views

88%

The Most Uncomfortable Question in Health Tech

What if the real problem in health care is not that we lack enough intelligence, but that we keep mistaking activity for value?

That question becomes impossible to ignore when you place two facts side by side. On one hand, governments and health systems are racing to adopt artificial intelligence, promising earlier diagnoses, better triage, lower costs, and more personalized care. On the other hand, some of the best randomized evidence in medicine keeps showing a sobering pattern: giving people more medical services often does not make them healthier in any meaningful aggregate sense.

That tension matters because AI is entering a system already vulnerable to a classic illusion. Health care has long rewarded interventions that are visible, measurable, and easy to expand. More scans, more visits, more alerts, more risk scores, more claims of precision. But if the marginal unit of care often delivers little or no benefit, then AI can become not a cure for medicine's inefficiency, but an accelerant for it.

The deeper issue is not whether AI works in a narrow technical sense. It is whether we know how to govern intelligence when the system it inhabits is already biased toward overproduction, false confidence, and uneven benefit. In other words, the central challenge is not just building smarter tools. It is deciding what kind of health system deserves to be made smarter.

The Productivity Trap in Medicine

Medicine has a peculiar status. Few domains enjoy so much trust while being so difficult to evaluate at the level that matters most: outcomes for real populations over time. A treatment can improve a biomarker, increase visits, or boost adherence metrics, yet still fail to improve the thing patients care about, which is whether they live longer, feel better, and suffer less.

That is why the most revealing studies are often not the ones that track a single disease endpoint, but the ones that ask a harder question: if we broadly expand access to medical services, does population health actually rise? Repeatedly, the answer has been underwhelming. More insurance often increases utilization, but utilization is not the same as health. More medical contact can uncover more problems, generate more procedures, and make the system look busier without making people better.

This is the first mental model worth keeping in view: medicine has a throughput problem disguised as a value problem. Systems that measure input volume more easily than outcome quality are prone to confusing motion with progress. A hospital can look more active, a clinic can look more connected, and an insurer can look more generous, while the net health effect remains flat.

Consider a factory that proudly reports the number of gadgets assembled per hour but never checks whether the gadgets actually work. If a new machine increases production speed, leaders may celebrate, even if it also increases defects. Health care often behaves this way. We count interventions because they are easy to count. We struggle to count avoided harm, unnecessary anxiety, or treatments that never should have happened in the first place.

That is the uncomfortable backdrop for AI. A system that already overvalues activity will not automatically become wiser when equipped with prediction engines. It may simply become more efficient at producing more of the same.

Why AI Magnifies Old Mistakes

AI in health is often introduced with the language of rescue: it will reduce clinician burden, catch disease earlier, and extend expertise to more people. Those promises are plausible. But AI also has a hidden property that makes it unusually powerful and unusually risky: it scales decisions, not just labor.

A traditional clinical error is local. One doctor overorders tests, one clinic over-triages, one administrator adopts a poor workflow. AI can replicate those judgments instantly across thousands or millions of cases. That means a small calibration error, a biased training set, or a poorly chosen objective can turn into a systemwide pattern of harm or waste.

This is why governance is not an administrative afterthought. It is the steering mechanism that determines whether intelligence serves health or merely serves the appearance of health. If a model is optimized to maximize revenue, visits, or flagged risks, it may do exactly that while leaving outcomes unchanged. If a model is trained on biased data, it may preserve inequities while appearing neutral. If a model is deployed without feedback loops, no one may notice its failures until they become institutionalized.

AI does not enter a blank slate. It enters a preexisting moral and economic ecosystem, and then it amplifies whatever that ecosystem already rewards.

That insight links the ethics of AI directly to the evidence on medical inefficiency. The question is not simply, “Can the algorithm predict?” The deeper question is, “Predict for what purpose, under what incentives, and with what accountability?” A brilliant model in a badly governed system can increase harm faster than a mediocre model in a cautious one.

There is a temptation to assume that because AI is data driven, it will be more objective than human decision making. But objectivity is not a property of data alone. It depends on what counts as success. An AI model can be perfectly accurate at predicting who will be admitted, billed, or scanned, and still be ethically misaligned if those predictions encourage low value care or worsen disparities. In medicine, the target matters as much as the tool.

The Real Unit of Innovation Is Not the Model, but the Decision

Most debates about AI in health focus on model performance: sensitivity, specificity, area under the curve, accuracy, calibration. Those are important, but they are only halfway to the real question. The decisive unit is not the algorithm. It is the decision pathway into which the algorithm is inserted.

Think about an AI system that identifies patients at risk of deterioration. If the hospital lacks enough nurses, no rapid response protocol, and no evidence based treatment pathway, then the model may merely create anxiety and more documentation. The prediction is real. The benefit is not. By contrast, a simpler model embedded in a strong workflow, paired with clear escalation rules and human oversight, may save lives.

This leads to a second mental model: AI is not a product, it is a policy choice encoded in software. Every model bakes in assumptions about whose risk matters, what action should follow, and who bears responsibility when things go wrong. That is why the best governance frameworks emphasize accountability, transparency, inclusion, and responsiveness to end users. These are not abstract ideals. They are the practical conditions under which intelligence can become beneficial rather than extractive.

There is also a deeper lesson here about clinical culture. The health system often treats more information as inherently better, yet information without action is just noise, and action without evidence is just momentum. AI can produce more alerts than humans can absorb. It can uncover marginal risks that are statistically interesting but clinically irrelevant. Unless governance defines the threshold at which information becomes useful, AI may flood clinicians with urgency and drain attention from what truly matters.

The most advanced health systems will not be those with the most models. They will be those with the strongest decision hygiene. That means asking, before deployment:

What is the decision this model changes?
What outcome does that decision improve?
What harms might increase instead?
Who gets to contest the output?
How will we know if the model is helping after it goes live?

These are governance questions, not just technical questions. And they matter precisely because medicine has so often mistaken more intervention for better care.

A Better Framework: From Intelligence to Stewardship

The most useful way to connect these ideas is to shift the goal from maximizing intelligence to maximizing stewardship.

Stewardship asks a different set of questions than optimization. Optimization asks how to make a system perform better according to a chosen metric. Stewardship asks whether the metric itself is worthy, whether the gains are fairly distributed, and whether the system preserves trust over time. In a health context, stewardship means treating AI as a public trust technology, not merely a private efficiency tool.

This matters because health is not just another market good. Patients are not ordinary consumers. They are vulnerable, often under informed, and frequently unable to evaluate the quality of care in the moment it is delivered. That asymmetry makes accountability central. A system that can quietly extract value from patients by increasing procedures, nudging utilization, or automating dubious recommendations is not just inefficient. It is morally unstable.

If the best medical evidence repeatedly warns that more care does not automatically improve outcomes, then AI governance must begin with humility. The default assumption should not be that a more sophisticated system will naturally produce better health. The default assumption should be that any new tool may intensify existing distortions unless its incentives are carefully aligned with genuine benefit.

This is where public benefit becomes more than a slogan. Public benefit means the technology must work not only for early adopters, wealthy hospitals, or digitally legible patients, but across countries, populations, and care settings. It means acknowledging that the same algorithm can have different consequences depending on whether a clinic has staff to act on its output, whether a patient can afford follow up, and whether the surrounding system values prevention over billing.

In that sense, ethics is not a constraint on innovation. It is the only way to tell whether innovation is real.

The Practical Test: Does the System Get Better or Just Busier?

A useful way to judge any AI deployment in health is to ask a deceptively simple question: does it make the system better or just busier?

Better means fewer avoidable deaths, less suffering, reduced inequity, more trustworthy decisions, and better use of scarce clinical attention. Busier means more clicks, more referrals, more alerts, more billing, more data, and more apparent sophistication. The two can look similar from the inside, especially in the early stages of adoption.

Here is why that distinction matters. Many technologies initially appear successful because they increase measurable activity. A predictive model that sends more patients to the hospital may look useful if the hospital measures admissions rather than downstream outcomes. A triage tool that flags more risk may look proactive if nobody audits false positives. A diagnostic assistant may look impressive if clinicians feel supported, even when the system has simply shifted workload elsewhere.

That is why evaluation must be longitudinal and adversarial in the best sense. Health AI should be tested not only against human performance but against counterfactuals that capture real system effects. Did the model reduce overall harm, or did it just change where the harm appears? Did it improve access, or mainly increase demand for already scarce services? Did it narrow disparities, or only improve outcomes for people already easiest to serve?

A strong governance system would insist on post deployment monitoring, independent auditing, and the ability to retire models that fail in practice. This is especially important because AI systems can drift as populations change. Unlike a static guideline, a model can slowly become wrong while remaining operationally impressive. Governance is the mechanism that keeps a powerful tool from becoming an invisible liability.

Key Takeaways

Do not confuse more care with better care. Before celebrating any health innovation, ask whether it improves outcomes or merely increases utilization.
Treat AI as a decision system, not a gadget. Its value depends on the workflow, incentives, and human actions surrounding it.
Measure what matters, not what is easy. Track health outcomes, equity, false positives, downstream burden, and patient trust, not just accuracy metrics.
Build governance into deployment. Require accountability, auditing, contestability, and post launch monitoring from the start.
Use stewardship as the north star. The goal is not to maximize technical sophistication, but to produce durable, fair, and genuinely beneficial health gains.

The Future of Health AI Is a Test of Moral Clarity

The promise of AI in health is not wrong. It is incomplete. Intelligence can help medicine, but only if medicine first becomes honest about its own limits. If a health system already confuses activity with value, then AI may become the ultimate instrument of that confusion, faster, cleaner, and harder to challenge.

But if AI is governed as a public good, with clear accountability and a relentless focus on outcomes, it could do something more important than automate care. It could expose how much of modern medicine is driven by habit, incentives, and inertia rather than by actual healing. In that sense, AI is not only a tool for diagnosis or prediction. It is a diagnostic tool for the health system itself.

The real question, then, is not whether machines will become better at medicine. It is whether we will become better stewards of medicine before we ask machines to scale it.

Because if more medicine has not reliably meant more health, the next leap forward will not come from intelligence alone. It will come from the wisdom to use intelligence less naively.