Health first. AI second.

The fear that AI would replace healthcare and public health workers was the loudest story when this conversation started. It made for good headlines and it sold a lot of consulting work. It never really held up — not because the technology is weak, but because the framing was wrong. We are not in the business of removing humans from health. We are in the business of keeping them in the loop, especially where decisions touch bodies, families, and communities.

The right framing is augmentation. AI does not take the job; it changes what the job is. When a machine carries the repetitive load — coding, billing, scheduling, transcription, prior authorization — the worker who was doing those things gets back something every health system claims to value and almost never returns: time. Time for the patient. Time for the community. Time to actually supervise the decisions the model is now nudging in the background.

Before we talk about augmenting anything with AI, though, we should be honest about the order of operations. For more than a century, public health has been doing something specific. Determinants of health. Epidemiology. Health behavior. Policy design. The core functions that allow a population to be assessed, protected, and held in mind. None of that came from a model. It came from people walking communities, reading death records, asking inconvenient questions and sometimes losing their funding for asking them. That body of knowledge is the substrate of our practice. It is not the warm-up act for a more important technical conversation. Any AI/ML method that arrives without fluency in those domains will reproduce the same harms we have spent generations learning to name.

So we start there. Health first. AI second. The order matters and the order is the argument.

What AI does change is the determinants themselves, and this is the part of the conversation public health has not yet fully claimed. Broadband access is now a determinant of health. Digital literacy is a determinant of health. The information environment a person swims in — what gets amplified, what gets buried, what an algorithm decides to surface at the exact moment someone is choosing between a vaccine and a rumour — is a determinant of health. So is whether the model that just denied a prior authorization was ever audited, and by whom. So is who owns the data your patients produced for free. We can call these digital determinants of health[1], or information determinants[2], or whatever language survives the next conference cycle. The point is the same. They belong to us. They are not the engineering team’s problem to handle quietly while the rest of us look away.

This is why I keep saying that our role in this conversation is not to be AI users. It is to be AI critical thinkers. The difference is the difference between operating a tool and being able to interrogate it. A critical thinker can ask what the model was trained on, who it was tested on, where it fails, and who pays the cost when it fails. A user can press the button. We need more of the first kind, and our schools are not producing them fast enough.

In 2026, a serious public health curriculum has to do four things at once. It has to teach the foundations of AI/ML to people who are not engineers but will work alongside engineered systems for the rest of their careers. It has to teach design thinking, because most failures I see in deployed systems are not failures of math — they are failures of framing. It has to teach AI ethics, and not as a single elective in the last semester. And it has to be taught by people who actually practice public health. You cannot teach the why behind the what if you have never had to do the what yourself.

For practitioners already in the field, the bar is lower but the urgency is the same. Minimum viable AI literacy is the ability to use the common tools, evaluate their outputs, and know when to trust and when to question. That is the port of entry. After that comes fluency — the level that lets you navigate the new paradigm rather than just operate inside it. We are nowhere near a workforce with that fluency, and we keep treating that as somebody else’s planning problem.

A word about the workforce the consulting decks call “exposed.” Coding, billing, scheduling, prior authorization, transcription. The honest answer is that yes, these roles are exposed, and pretending otherwise does not help the people in them. The performance of current AI on these tasks is real. What we owe those workers is not denial; it is a retraining path that lands them somewhere useful, and somewhere they actually want to be. The Digital Public Health argument has a useful staircase here[3]. Digitization is moving an analogue process into digital form. Digitalization is integrating digital tools into how the work already gets done. Digital transformation is redesigning the workflow around the tools. Each step opens new roles — digital epidemiologists, AI health informaticians, health systems designers, public health AI auditors, health data stewards. None of those job titles were on a syllabus ten years ago. All of them need to be on one this year.

A word also about radiology, because the replacement conversation always comes back to it. The prediction that AI would replace radiologists was confident, loud, and wrong. It came back every season for years and every season the people doing the actual work pointed out the parts it had ignored. The performance numbers on image classification keep getting better. So do the reports of hallucination and misclassification when those same models meet the real world. Human-in-the-loop has become more important in radiology, not less. The lesson generalizes. Lack of adoption is not function failure. Adoption without a deployment plan is not progress. The interesting debate is not whether the model scores better than a junior resident on a benchmark. It is whether we can deploy it at scale, maintain it under drift, and protect patients and communities while we do.

If we are still doing this work in 2035 — and we should be — the sentence I want to be able to write is small. AI expanded the human factors in the practice of health. It did not replace them. It did not erase them. It gave them back the room they had been losing.