How Public Health Agencies Can Use Artificial Intelligence Effectively and Responsibly

Courtesy of Andrea King

My career has been entirely in the public sector. I started in academia and trained as an epidemiologist, statistician, and data scientist, with a background in international affairs. I worked for the New York City Department of Health and Mental Hygiene before moving to the Washington State Department of Health. Over time, I went from working with an individual dataset to trying to improve the systems that produce and connect data across agencies and jurisdictions. 

When the newer generative AI tools emerged, it was clear they could change how this work gets done. But with accelerated technology come questions about how public health agencies can adopt these tools responsibly. I believe there’s a significant gap between what public health practitioners think you can do with AI and its actual capabilities, which are far beyond the field’s general knowledge, and my goal is to close it by showing what AI can do.

PubHealthAI is a practitioner-led network focused on the responsible adoption of AI for public health. It gives professionals from state, local, territorial, and Tribal governments, as well as academic institutions, nonprofits, and other mission-aligned organizations, a space to compare experiences, share knowledge, and highlight good work. Because the technology is moving so quickly, there is real value in learning together rather than having every agency struggle on its own.

The network began in June 2024 with a small group in the Pacific Northwest. I initially convened it to share results from a proof-of-concept project that used AI to turn unstructured text from Washington’s overdose reporting system into structured data for analysis. As I started sharing more results, I noticed interest growing, and now we have about 700 people on our distribution list.

The network has become a place to exchange approaches, tools, and evaluations across institutions and, increasingly, countries. We have members in Australia, Singapore, Hong Kong, several countries in Africa, and beyond. We’ve noticed that many countries are facing the same fundamental challenges on how to approach AI: governance, privacy, security, and organizational capacity.

The early benefits are often operational: saving time, improving accuracy, lowering costs, and expanding staff capacity. Many implementations are recent, so it may take longer to measure their downstream effects on health outcomes.

Some of the strongest use cases involve processing huge volumes of information that would be impractical or impossible for people to handle manually. A lot of public health departments still use paper records, and AI can help make these records useful instead of having them sitting somewhere in storage boxes.

For example, one state health department used AI-supported document extraction as part of an effort to digitize 100 years’ worth of paper vital records, or about 50 million records. No human can sit and transcribe all of those. This streamlining of time and resources helped the state save an estimated $300,000 over five years. In resource-constrained environments, that’s huge.

Other practical uses include coding support, data cleaning, visualization, troubleshooting, and building and monitoring data pipelines. The technology is incredibly effective, but staff still need to have enough technical expertise to read, understand, and validate the AI output before it is used.

Institutional culture can be a major barrier. Risk-averse, consensus-driven organizations may spend so long trying to write a perfect policy that the technology has changed by the time the policy is finished. Meanwhile, employees may begin experimenting on their own with consumer tools that the agency cannot necessarily see or manage, and that opens up privacy concerns. It’s generally more useful to put a framework and an approved tool in place, then update both as conditions change. We can’t let perfect be the enemy of the good.

Procurement can be another barrier. For a lot of governments and organizations, contracts for purchasing new tools can take a year or more to execute. But since the technology is changing all the time, finding opportunities to shorten these procurement timelines is important.

Other obstacles include inflexible technology environments, limited staff time for learning, and short-lived pilots that lack sustained support.

Agencies should start with a problem they need to solve, then assess whether AI is an appropriate part of the solution. Simply buying licenses and telling staff to experiment is not enough. Clear, concise governance and policy frameworks, organization-wide training, and leadership fluency are important. Executives need to understand the technology well enough to make informed decisions, and employees need support to develop and apply new skills.

Agencies shouldn’t treat AI as completely separate from every other technology. Existing confidentiality, security, and governance standards should still apply. The goal is to add AI-specific requirements where they are needed, not to rebuild the entire policy structure from scratch.

AI tools need to be routinely reviewed, even after deployment, to make sure they are functioning the way agencies think they are. Staff also need to understand how training data can carry existing human biases and how to test for and mitigate those effects.

For instance, most AI models are largely trained on just a fraction of the languages spoken globally, relying heavily on English-language content for their knowledge base and internal operations. Thus, they contain biases inherent in English-speaking countries, and performance varies if they’re prompted in languages not common in their training base. This means that if an agency is, say, deploying a public-facing chatbot with multilanguage capabilities, it’s important to test thoroughly, both with inputs and outputs in non-English languages, and periodically evaluate performance. While AI models continue to improve on this front, I’d also suggest adding a “Machine translated by ‘name of the tool,’ which may make mistakes” disclaimer so users are aware.

Agencies should also use a centralized, enterprise-level platform rather than having employees use personal accounts for official business. That centralization makes oversight, records management, data privacy, and compliance much easier.

In an ideal world, AI would become a routine layer within public health systems, as easy to use as the internet. It could support surveillance systems, registries, case management, and the back-office work needed to connect older technologies. Rather than expecting agencies to replace entire technology stacks, AI tools would work across different systems and help standardize, clean, and move data among jurisdictions.

That could free public health professionals from spending enormous amounts of time aligning data and allow them to focus on more meaningful activities, such as turning analyses into recommendations and working within communities. AI literacy would also become a core workforce competency, with experienced staff feeling as comfortable adapting to new tools as employees entering the field fresh out of school.

Know where your expertise matters most and keep developing it. Public health depends on knowledge of the field, human behavior, communities, and the practical work of protecting population health. Rather than defining a job around a single task, such as cleaning a dataset or writing code, people should think more broadly about the value they bring and how technology can help them extend it.

The goal is not to become less of an expert because AI can perform certain tasks. It is to become a stronger public health practitioner and use the technology in service of that expertise.