Public health has always been an argument about what a society owes the people in it.
The argument has been settled differently in different centuries. It was settled one way when cities built sewers, another way when vaccination became compulsory, another way again when tobacco advertising was restricted. Each time, someone had to decide that a collective action was warranted, find the authority to take it, and answer for the consequences.
The instruments changed. The structure of the decision did not.
Artificial intelligence is the newest instrument. It is a genuinely powerful one, and this book takes that seriously. But it enters a field where the binding question has never been what could be known. It has been what could be done about what was known, by whom, with whose permission, at whose expense.
What the Instrument Actually Offers
Begin with an honest account of the capability, because the argument that follows is weaker if the capability is understated.
Artificial intelligence can identify patterns in volumes of data no human team could read. It can find signals in text, images, and sensor readings that were previously inaccessible at scale. It can flag an anomaly in reported symptoms across thousands of clinics faster than a surveillance system built on weekly reporting. It can read a retinal image or a chest radiograph with accuracy comparable to a trained specialist, in places where no trained specialist is available. It can identify which households in a city are most likely to contain a hazard, which patients are most likely to deteriorate, which communities are least likely to be reached by a standard outreach campaign.
These are real capabilities and they are not trivial. Public health has spent two centuries trying to see earlier and reach further. Here is a tool that does both.
The Question the Tool Cannot Answer
Now consider what happens the moment any of that succeeds.
A surveillance system flags an unusual cluster of respiratory presentations in a district. The signal is correct. What follows depends on whether a public health authority exists in that district with the legal standing to investigate, the staff to deploy, the budget to sustain a response, and the political cover to act on a signal that might turn out to be nothing.
A screening program identifies a person with early disease in a community with no specialist within two hundred miles. The detection is correct. Whether it becomes a treated case depends on referral pathways, transport, cost, and whether anyone follows up.
A predictive model identifies the households in a city most likely to contain lead hazards. The prediction is accurate. Whether a single child is protected depends on whether the city has inspection authority, remediation funding, and the ability to compel a landlord.
In each case the technology performed. In each case the outcome turned on something the technology does not supply and cannot supply.
That is the argument of this book, stated plainly: Artificial intelligence can help public health see sooner. It cannot, by itself, give institutions the authority, resources, legitimacy, or accountability required to act on what it reveals.
Everything that follows is an examination of that gap. Where it opens. What widens it. What closes it.
Why This Is Not a Familiar Complaint
There is a common version of this observation that amounts to saying technology is not a magic solution. That version is true and useless.
The argument here is more specific. It is that prediction and prevention are separated by an institutional structure, and that structure can be examined, described, and in many cases changed. The gap is not a lament. It is a set of conditions.
Some of those conditions are legal. An agency either has the statutory authority to act on a finding or it does not, and that is knowable in advance.
Some are financial. A detection program without a treatment budget produces a longer list of untreated people, which is not a neutral outcome.
Some are procedural. A model that produces an alert nobody is assigned to review has generated a record of institutional knowledge without producing institutional action, and that record exists whether or not anyone acted on it.
Some are contractual. The terms under which a health system procures an artificial intelligence tool determine what it can inspect, what it can demand, and what it can prove afterward. Those terms may be negotiated far from the teams that will eventually use, evaluate, or oversee the system. Yet they can determine which information the institution receives, which obligations the supplier accepts, and what evidence exists when something goes wrong.
And some are about legitimacy. A community that does not trust the institution holding the prediction will not act on it, will not participate in the program built around it, and will be right to ask why its data was collected in the first place.
None of these conditions are technical. All of them determine whether the technology produces health.
Why AI Succeeds and Fails
There is a pattern in the record, and it is not the pattern most people expect.
Artificial intelligence in health does not fail only because a model is wrong. Some of its most consequential failures occur when the model performs as intended and the surrounding system cannot absorb, interpret, or act on the result.
It fails when a tool is validated on one population and deployed on another. It fails when an alert is added to a workflow already producing more alerts than anyone can review. It fails when a system is trained on a measurement that stands in for the thing anyone actually cares about, and the substitution turns out to carry a history. It fails when nobody outside the organization that built it can examine why it produced a given output. It fails when it is deployed into a setting without the electricity, connectivity, staffing, or follow-up capacity the deployment assumed.
And a particularly consequential failure occurs when an institution acquires a capability without establishing the authority, responsibility, and capacity required to use it.
The successes follow the inverse pattern. They tend to involve modest technical achievements paired with unusually well-prepared institutions. Someone had already established who would act on the output. Someone had already secured the budget for what happens after detection. Someone had already built the community relationships that make participation possible. The technology was the last thing added, not the first.
This is the same lesson public health learned about sanitation, vaccination, and every other instrument it has adopted. The technology is necessary. It is never sufficient. What determines the outcome is the institutional capacity built around it.
How to Read This Book
The chapters ahead are organized to make that pattern visible rather than merely assert it.
We begin with history, because artificial intelligence is not the first powerful instrument public health has had to absorb. Then we establish enough technical literacy to understand what artificial intelligence can and cannot reasonably be asked to do.
The applications follow, examined at full strength: surveillance, diagnostics, social determinants, mental health, information environments, and emergency response. Each describes what artificial intelligence genuinely contributes, and each eventually reaches the same point. Something has been revealed, and a decision now has to be made by someone with the standing and capacity to make it.
The failures come next, consolidated rather than scattered, because the mechanisms repeat across domains and are clearer examined together.
Then equity, which asks who benefits, who bears the error, and what happens when a prediction reproduces the conditions that produced it.
Then implementation, which asks how any of this survives contact with an operating health system.
Then the instruments that make expectations binding, which is the question health leaders are least prepared for and the one that determines whether everything before it matters.
Then the human element, which is not a sentimental postscript but the answer to a specific question. What should people remain responsible for precisely because these systems exist?
The book closes with working tools. They ask what your organization would do if a system worked, whether you could respond equitably to what it found, who would be authorized to act on it, and how you would know afterward whether any of it was working. They are meant to be completed rather than read.
What Reclaiming Means
The title of this book is not a slogan.
Health has been ceded, gradually and mostly without decision, to systems that were never designed to protect it. To commercial environments that shape what people eat and breathe. To information environments that shape what they believe. And now, increasingly, to algorithmic systems that make consequential determinations about people who have no way to see, question, or appeal them.
Institutions, meanwhile, can accumulate knowledge about risk faster than they accumulate the authority and capacity to respond to it.
Reclaiming health means taking those decisions back into view. Not rejecting the instruments, which would be both futile and wrong, but insisting that someone be visibly responsible for what is done with them.
That is not a technical project. It is the same project public health has always been engaged in, with a new instrument and a familiar question.
What are we prepared, and authorized, to do with what we know?