Artificial Intelligence in Pandemic Preparedness, What the Evidence Actually Supports, Curely AI

Share

Artificial Intelligence in Pandemic Preparedness, What the Evidence Actually Supports

Bottom line. AI is now a working layer in pandemic preparedness, not a future promise, but its value is concentrated and uneven. The strongest, best-evidenced gains are in early signal detection and genomic surveillance. The weakest and most over-promised are long-range forecasting and autonomous prediction. For health systems and public health agencies deciding where to put scarce resources, the honest read is that AI earns its place when it augments human-led surveillance and sharpens response, not when it is asked to replace epidemiological judgment.

Where the evidence is strongest, detecting signals early

The clearest, most durable wins are in epidemic intelligence, the work of catching an unusual signal before it becomes an official case count. AI systems that scan news, official reports, and social media have a documented record of surfacing outbreak signals ahead of formal channels. BlueDot is the most cited example, having flagged the emergence of COVID-19 in Wuhan in late 2019 before the wider global response took shape. Platforms such as HealthMap and EPIWATCH apply the same approach, mining open digital sources for early anomalies that warrant human review.

This is now institutional infrastructure rather than experiment. The WHO Hub for Pandemic and Epidemic Intelligence in Berlin worked with more than 160 Member States and over 190 partners in 2025 to build detection systems and shared tooling. The scale of the signal problem is the reason AI is involved at all. WHO has reported identifying roughly 4,500 signals of potential public health events every month, a volume no purely manual process can triage. At its 2026 Annual Meeting, the World Economic Forum announced two AI-enabled platforms intended as global public goods, a Pandemic Preparedness Engine and a Global Pathogen Analysis Platform, aimed at faster identification and characterisation of threats across countries.

Evidence grade. Strong on documented case studies and growing institutional adoption. More mixed on rigorous, head-to-head effectiveness, since much of the proof remains retrospective and system-specific rather than drawn from prospective comparison. The capability is real. The independent measurement of how much earlier and how much more accurately these systems detect outbreaks is still maturing.

The prediction trap, what the failures taught us

Preparedness narratives often blur detection and prediction, and the distinction is where AI has stumbled. Detection asks what is happening now. Prediction asks what will happen next, weeks or months out, and that is far harder.

The cautionary case is Google Flu Trends, which initially estimated influenza prevalence from search patterns but later overestimated flu levels badly. The failure was not bad luck. It came from overfitting and from a model that did not account for media-driven changes in how people search during a scare. The deeper lesson was argued directly in Nature by Holmes, Rambaut, and Andersen, who urged funders to spend on surveillance rather than prediction, on the grounds that reliable observation beats speculative forecasting when stakes are high.

Evidence grade. Strong and well-established. The limits of long-range epidemic forecasting are documented across multiple independent analyses. The practical implication is concrete. Treat AI forecasts as decision support with explicit uncertainty, not as a substitute for the surveillance that grounds them.

Surveillance under constraint, the African frontier

The gap between capability and deployed infrastructure is widest in low-resource settings, which is also where the next outbreak is as likely to begin as anywhere. The 2024 mpox emergency made this concrete. WHO declared a Public Health Emergency of International Concern in August 2024, with first confirmed cases reported in Burundi, Kenya, Rwanda, and Uganda. The question was not whether AI could help in principle, but whether it could work where laboratory capacity, connectivity, and trained personnel are limited.

Several efforts are testing exactly that. AI4Mpox, led by Canada's International Development Research Centre and the University of Toronto since 2025, runs across seven mpox-affected African countries and pairs AI and modelling with real-time dashboards and community-level data for hotspot detection. In Sierra Leone, teams working with the Institut Pasteur de Dakar deployed decentralised genomic sequencing and predictive modelling during one of West Africa's largest recorded mpox outbreaks in 2025, a model of real-time response built for constrained conditions. Edge AI and AI-enabled point-of-care diagnostics are the technical bridge, allowing local analysis where centralised infrastructure is unavailable.

The frontier also shows the evidence problem in sharp relief. On-device, offline mpox image screening tools have reported high accuracy, in one case around 96 percent, which is genuinely promising for triage where laboratory access is scarce.

Evidence grade. Emerging and regionally specific for the deployed programmes, which are recent and still being evaluated. Weaker for the standalone screening tools, which often rest on single studies, sometimes with vendor affiliations, and have not yet been independently validated at scale. Promising is the correct word. Proven is not, yet.

The real bottleneck, from signal to action

The honest constraint in preparedness is not algorithmic. AI epidemic intelligence systems remain, in the words of recent reviews, fragmented and reactive. They struggle with filtering misinformation, integrating across data sources, and adapting in real time as an outbreak evolves. Many are built for either detection or response, but not both, and they do not update fluidly as conditions change.

Two further frictions are structural rather than technical. Data protection regimes such as GDPR and HIPAA, designed for good reason, constrain the real-time exchange of health data that cross-border detection depends on. And the consensus among the public health intelligence community, including voices at the WHO Hub, is that keeping humans in the loop remains essential for responsible use. The value of an earlier signal is only realised if a health system trusts it, can act on it, and has the surge capacity to respond. That is a workflow and governance problem, not a modelling one.

What this means for preparedness

The useful posture is neither dismissal nor hype. AI has earned a durable role in the detection and surveillance layer of pandemic preparedness, and a credible, fast-developing one in genomic and field response in constrained settings. It has not earned trust as a long-range oracle, and it does not remove the need for laboratory capacity, trained epidemiologists, clear data governance, and the surge infrastructure that turns a signal into a response.

For decision-makers, that suggests a specific investment order. Strengthen the surveillance backbone first, then layer AI where it measurably shortens the distance from signal to action, and hold forecasting claims to a higher evidentiary bar than detection claims. The systems that will matter in the next outbreak are the ones built around real public health workflows, on good data, with humans deciding. The intelligence layer is real. Its job is to make good responders faster, not to replace them.

Artificial Intelligence in Pandemic Preparedness, What the Evidence Actually Supports