Predicting pandemics, predicting people: the ethical limits of AI-powered public health surveillance

Review

Adam and colleagues examine the ethical boundary between AI systems that improve epidemic intelligence and systems that increasingly infer, classify, and act upon individuals or groups. The mini-review's central claim is that predictive capability is not a sufficient basis for legitimacy. As surveillance moves from detecting population-level disease patterns toward behavioral inference, profiling, and intervention, governance must establish necessity, proportionality, purpose limitation, data minimization, equity, transparency, meaningful human oversight, accountability, and redress.

The paper's most useful contribution is its treatment of surveillance as a continuum of institutional power rather than a binary question of whether AI is present. The shift from observation to inference, prediction, classification, and intervention progressively changes what public-health authorities can know and do. This makes the relevant governance question not simply whether a model is accurate, but whether the resulting action remains justified by a specific public-health purpose and bounded by rights. The paper is particularly clear that individual-level prediction is not inherently illegitimate. Targeted vaccination, exposure notification, and contact tracing may require individual risk assessment. The boundary instead turns on mandate, necessity, proportionality, and safeguards against repurposing.

A second contribution is the insistence that equity is a performance property, not an ethical supplement. Missing data, uneven digital access, poor local fit, and transferred models can distribute false positives and false negatives unevenly, especially in resource-constrained settings. This connects technical validation directly to institutional legitimacy: a model with acceptable aggregate accuracy may still be unfit for public use if its errors systematically burden particular groups.

The lifecycle framing is also governance-relevant. The paper calls for pre-deployment assessment of legal authority, data provenance, subgroup performance, and local applicability; operational requirements for human override and correction; and post-deployment monitoring of drift, incidents, workflow effects, and unexpected harms. It also includes suspension, rollback, replacement, and retirement as legitimate governance outcomes. That moves beyond the common assumption that oversight means approving deployment and then monitoring performance indefinitely.

The evidentiary limitation is that this is a mini-review rather than a systematic review or empirical evaluation. It synthesizes recent literature into a normative framework, but does not demonstrate that the proposed principles have been operationalized successfully in real public-health institutions. Terms such as meaningful human oversight, community participation, proportionality, and public trust remain institutionally under-specified. The paper identifies the need for accountable institutions but gives less detail on who may authorize deployment, who can suspend a system, how affected individuals gain standing to challenge predictions, what evidence must be preserved for appeal, or how conflicts between public-health urgency and rights constraints are adjudicated.

The strongest governance implication therefore lies in the paper's concept of an ethical ceiling. Surveillance capacity should not become permanent infrastructure merely because it has demonstrated utility. A system may remain technically effective while losing legal, social, or ethical justification as threats, data conditions, public expectations, or institutional capacity change. Operational governance must therefore include revocable authority and explicit exit conditions, not only model monitoring.

The paper establishes a coherent rights-preserving framework for judging AI surveillance proposals. What it does not yet establish is an executable institutional architecture for enforcing those limits. The next useful step would be to translate necessity, proportionality, contestability, redress, and retirement into named decision rights, evidence requirements, thresholds, and accountable authorities that can be tested across real surveillance programmes.

Key Insight

The legitimate boundary of AI public-health surveillance is not the point at which prediction becomes technically possible, but the point at which institutions can no longer justify, constrain, contest, and retire the resulting power over people.