Table of Contents
Published: July 20, 2026
Read Time: 3.9 Mins
Total Views: 7
Every argument for bringing AI into public health eventually arrives at the same challenge: trust. Public health has limited power to compel. Rather, it works by persuasion, and persuasion depends on a public that believes the institution is competent, honest, and acting in its interest. AI can strengthen that relationship or corrode it, and this depends on how AI systems for public health are built, deployed, and governed.
Extend the workforce, or hollow it out
A large share of public health work is what might be called cognitive manual labor: the repetitive, rule-based handling of data and documents that consumes the hours of skilled people. AI is well designed to absorb that labor, which creates a fork in the road. Used to extend the workforce, it can free epidemiologists from cleaning spreadsheets and move them into the field, where the human work of engaging communities and building relationships actually happens. Used to justify cutting headcount, it removes the very people who carry an agency’s credibility in the neighborhoods it serves. The pandemic made unmistakable how much public health depends on trusted humans, and an agency that automates those humans away will find that no model can rebuild the relationships it dissolved. A related hazard is skill atrophy, the slow loss of the manual competence to recognize when an automated system is wrong, which is why workflows should require human verification and why agencies should actively preserve analytic and communication skills rather than let them lapse.
The accuracy and equity problem is not hypothetical
AI introduces specific threats to trust, beginning with models that hallucinate, perform inconsistently across populations, and absorb the biases already embedded in our data. The hallucination risk is concrete: a researcher recently invented a fictitious skin condition, and within weeks chatbots were confidently telling people it was real, which is exactly the failure mode public health cannot afford when it speaks to the public during an emergency. The equity problem compounds it. Public health datasets often underrepresent marginalized communities and reflect unequal access to care, so a system trained without deliberate mitigation will reproduce those inequities at scale, and a tool validated only in well-resourced settings could widen the disparities public health exists to close. The same caution applies to the consumer-facing AI products marketed to individuals, where I have written about how companion chatbots have already become a public health problem.
What responsible deployment requires
Governance offers the answer, and it does not have to be invented from scratch. Established frameworks already exist, including the NIST AI Risk Management Framework, the World Health Organization’s guidance on the ethics and governance of AI for health, the CDC’s public health AI strategy, and emerging standards for the responsible use of AI in healthcare. Tools should document their data sources and known limitations transparently, be validated for accuracy before deployment, and be monitored afterward for drift, bias, safety, and unintended consequences. They must operate within the laws on privacy and data protection. They should be developed and tested in diverse populations, designed with input from the communities they serve, and made available through open licensing or subsidized access to the county and tribal agencies serving the highest-need populations and least able to pay. Above all, high-consequence decisions that affect access to services, outbreak measures, risk communication, or resource allocation should remain under meaningful human oversight, with transparency requirements scaled to the stakes.
The promise of AI in public health is genuine, and I have laid it out across surveillance, emergency response, chronic disease prevention, and climate threats. That promise will be realized only if the technology is built to serve the public rather than to economize on it, because a tool that erodes trust subtracts more than it adds, no matter how accurate its forecasts. Getting the technology right is the easy part. Getting the trust right is the work that determines whether any of it matters, and it is the standard against which every deployment should ultimately be measured.
