The Hard Part of AI in Public Health Is Not the Technology
- Autumn Henkle
- Apr 13
- 3 min read
The CDC’s recent AI strategy makes it clear that public health is moving past experimentation and into real implementation. The focus is no longer on what AI could do in theory, but on how it will actually be used to improve detection, data access, and decision making across systems. That shift is important, but it also brings a different set of challenges into focus. The potential of AI in this space is not new. What is less discussed is what determines whether that potential translates into something reliable in practice. At this point, the limiting factor is not the technology itself, but the systems it has to operate within.
Public health data environments are not designed as unified systems. They have developed over time across hospitals, laboratories, state agencies, and federal platforms, each operating with its own requirements and constraints. This means data is not just distributed, but inconsistent in structure, definition, and quality. Efforts to centralize or standardize this data often surface these inconsistencies rather than resolve them. Differences in reporting practices, variations in how data is defined, and gaps in completeness persist even after integration, creating a layer of misalignment that is difficult to fully eliminate. This is already a challenge on its own, but it becomes more significant once AI is introduced into the equation (Centers for Disease Control and Prevention [CDC], 2023).
AI systems depend on a level of consistency that these environments do not naturally provide. Models are designed to identify patterns across inputs that are assumed to be comparable and stable. When those assumptions do not hold, outputs become less reliable, not because the models are inherently flawed, but because the data reflects underlying inconsistencies across systems. In practice, this means that what looks like a modeling issue is often a system issue. Variability in how data is collected, delays in reporting, and differences in interpretation all affect the quality of the output in ways that are not always immediately visible (World Health Organization [WHO], 2021).
The challenge extends beyond the data itself into how it is processed and moved through systems. Data pipelines have to account for variation across sources without introducing additional distortion, while transformations must preserve meaning despite structural differences. At the same time, latency and completeness have to be managed in ways that align with how the data is actually used in decision making. These are not edge cases. They are ongoing conditions that shape how reliable an AI system can be in production. Without addressing them directly, improvements in model design have limited impact on overall performance.
Governance adds another layer that cannot be separated from the technical side. Public health data is subject to strict privacy, security, and compliance requirements, which influence how data can be accessed and used. Designing systems that support advanced analytics while remaining within these constraints requires coordination between architecture and policy. The CDC’s emphasis on interoperability reflects this reality. Interoperability is not just about enabling systems to exchange data, but about ensuring that the meaning of that data is preserved across contexts. Without that, aggregation at scale can introduce more confusion rather than clarity (Office of the National Coordinator for Health Information Technology [ONC], 2023).
What tends to differentiate organizations that make progress is not the specific tools they choose, but how they approach the system as a whole. Rather than treating AI as a standalone capability, they focus on aligning the environment around it. That includes standardizing definitions where possible, designing pipelines that can handle variation, and creating feedback mechanisms that surface inconsistencies early. It also means recognizing that data quality is not something that can be fixed once and considered complete. It is an ongoing characteristic of the system that has to be managed as processes evolve.
This leads to a different way of thinking about modernization. It is less about implementing a new platform and more about continuously aligning how data is produced, interpreted, and used across systems. Technical decisions are made with an understanding of how these environments actually function, not just how they are expected to function. That distinction matters, because it determines whether new capabilities can operate reliably over time or whether they introduce additional friction.
AI in public health is already being implemented, and its role will continue to expand. The organizations that see meaningful impact will not necessarily be the ones that adopt it first, but the ones that invest in the systems that support it. At this stage, the technology itself is not the primary barrier. The question is whether the systems in place are ready to handle what AI requires.
References
Centers for Disease Control and Prevention. (2023). Data modernization initiative.
Office of the National Coordinator for Health Information Technology. (2023). Interoperability
standards advisory. https://www.healthit.gov/isa/
World Health Organization. (2021). Ethics and governance of artificial intelligence for health.
Comments