AI Is Already Watching for the Next Outbreak

Inside WHO’s Epidemic Intelligence from Open Sources — and the wider landscape of AI-assisted outbreak detection.

When we imagine artificial intelligence being used against the next pandemic, it is tempting to picture the future: an AI system notices an unusual disease cluster, recognises a novel pathogen and warns governments before conventional surveillance catches up. Some of that future is already being built — and the World Health Organization’s Epidemic Intelligence from Open Sources initiative, usually known as EIOS, is one of the clearest examples.

EIOS is not an autonomous AI epidemiologist, and it does not replace national surveillance, laboratories or public-health professionals. It is better understood as a technology-supported epidemic-intelligence environment designed to help trained users identify and assess potential public-health threats from the enormous volume of information appearing every day. Its underlying problem is simple: the first public indication of an outbreak does not always arrive through an official disease-notification system. It might appear in a local news report, a government website, an online discussion, a report of unexplained animal deaths, a reference to hospitals receiving unusual numbers of patients, an environmental event, a food-safety notice, or information published in a language that analysts elsewhere are not routinely monitoring. Individually, most of these signals mean very little. Collectively, they can sometimes reveal the beginning of something important. The challenge is finding them — and that is a problem of scale that machines are unusually well suited to.

Before surveillance comes intelligence

Formal surveillance remains essential to public health: laboratories produce confirmed results, clinicians report disease, governments operate statutory notification systems, and international mechanisms such as the International Health Regulations provide formal routes through which significant events are assessed and communicated. Yet formal surveillance is only one way of learning that something may be happening. Public-health organisations also practise epidemic intelligence: the systematic collection, analysis and interpretation of information that may indicate a potential threat. Open-source information can be particularly valuable here because it is produced continuously and at enormous scale — and the problem is precisely that volume. No team of analysts can manually read every potentially relevant public source, in every language, every day. This is where technology becomes indispensable.

What EIOS actually is

EIOS describes itself as a collaboration between public-health stakeholders around the world rather than simply a piece of WHO software. Led by WHO, it brings together new and existing initiatives, networks and systems into what it calls a unified all-hazards, One Health approach to the early detection, verification and assessment of public-health risks using open-source information. At its centre is the EIOS system, a web-based platform through which authorised public-health users can monitor large quantities of open-source information, which the system helps organise, filter and categorise so that analysts can concentrate on material potentially relevant to health threats.

The scale is significant. By the end of 2025, on WHO’s account, roughly 120 countries and 30 organisations and networks were using the EIOS system, and it is now hosted at the WHO Hub for Pandemic and Epidemic Intelligence in Berlin — established on 1 September 2021 with substantial German government funding to transform global surveillance and build a collaborative intelligence ecosystem. The important distinction runs through all of it: the machine helps find the signal; people decide what the signal means.

Epidemic intelligence did not begin with EIOS

One reason EIOS matters is that it consolidates a field with real heritage rather than inventing it. Canada’s Global Public Health Intelligence Network (GPHIN), developed by the Public Health Agency of Canada, has monitored open-source information — news wires, discussion groups and websites — across multiple languages since the late 1990s, and famously contributed to early awareness of SARS; it still supplies a substantial share of EIOS’s input. ProMED, run by the International Society for Infectious Diseases, is an expert-moderated reporting network operating around the clock with tens of thousands of subscribers across some 200 countries, and it posted one of the earliest public alerts about the cluster of pneumonia in Wuhan at the end of December 2019. HealthMap, based at Boston Children’s Hospital, automatically aggregates online reports into a real-time map of possible outbreaks. EIOS connects these systems and actors rather than replacing them — the value is integration.

Where AI comes in

EIOS sits within a broader transformation in epidemic intelligence driven by natural-language processing, machine learning, automated classification and related computational techniques. Machines are very good at the things humans find exhausting: scanning huge quantities of information, grouping material by topic, recognising terms and entities, filtering duplicates, working across many languages, prioritising according to defined criteria, and finding patterns within streams of text. None of that means a machine understands an outbreak the way an experienced epidemiologist does — but it can dramatically reduce the search problem. An analyst interested in unusual respiratory disease might otherwise need to search thousands of sources by hand; with an epidemic-intelligence platform, information is continuously collected and organised so that potentially relevant reports rise towards the analyst. AI here acts less like an autonomous disease detective and more like an extraordinarily fast research assistant.

A rumour is not a case

This is where the limitations become important. A newspaper report of unexplained illness does not prove an outbreak exists, and a social-media post certainly does not. Several articles may repeat the same inaccurate original; translations can alter meaning; terminology differs between countries; and a dramatic event may receive enormous coverage while a more important biological signal receives almost none. Open-source surveillance therefore has to distinguish between signal detection and verification. EIOS can help analysts find something worth investigating; public-health professionals still have to determine whether it is credible — checking official sources, contacting national authorities, examining epidemiological data, reviewing laboratory information or comparing the signal against other surveillance systems. AI can help answer what should we look at? It cannot safely answer, alone, what is actually happening?

Indicator-based and event-based surveillance

The distinction becomes clearer if we separate indicator-based from event-based surveillance. Indicator-based surveillance works with structured information — numbers of cases, laboratory diagnoses, hospital admissions, mortality, disease notifications — which is invaluable but often arrives through established reporting structures. Event-based surveillance looks for information about events that may pose a health risk, where the information can be less structured and less certain: a report of unexplained animal deaths, an unusual cluster of illness, a possible food-contamination event, an unexplained environmental hazard. The purpose is not to treat every report as fact; it is to identify events requiring assessment. EIOS is particularly valuable within this event-based world — the world where the earliest and least certain signals live.

The speed advantage — and where the bottleneck moves

The attraction is obvious. Formal surveillance contains delays: someone becomes ill, decides whether to seek care, is evaluated by a clinician, perhaps has a sample collected, a laboratory produces a result, the information enters surveillance, and a cluster becomes visible. Open-source information can sometimes appear somewhere along that pathway before the formal system has assembled the complete picture. That does not make it more reliable than laboratory surveillance — it makes it potentially earlier, and for health security, earlier imperfect information can be extremely valuable if it triggers appropriate verification. It is the same tension that runs through all of our work: speed and certainty do not arrive at the same time.

We have described detection latency as the time between a biological event beginning and the surveillance system recognising a meaningful signal, and AI-assisted epidemic intelligence has genuine potential to reduce it: a relevant report published overnight in another language no longer has to wait for an analyst to stumble upon it. But reducing detection latency creates pressure elsewhere. If a system produces signals faster than institutions can assess them, the bottleneck simply moves — detection becomes faster, and interpretation becomes the problem. Suppose an AI-assisted system surfaces 200 potentially interesting signals in a day; ten look unusual, three appear credible, and one may indicate a serious emerging event. Which deserves attention first? That question makes human expertise more important, not less. An effective system should therefore not simply maximise the number of things detected; it should improve the ratio between useful signal and analytical burden — a very different objective from collecting everything.

The One Health opportunity

EIOS becomes especially interesting through a One Health lens, and notably it describes itself in exactly those terms. An emerging human-health threat may first become visible outside human medicine: a local report of unexplained poultry deaths, a veterinary announcement of unusual livestock disease, fish mortality, an invasive mosquito detected somewhere new, a food-safety notice, wildlife deaths in local reporting, or sewage contamination after heavy rainfall. None of these is necessarily a human outbreak, and some may never become one — but each can provide context for biological risk. Open-source epidemic intelligence can therefore extend beyond searching for reports of sick people; it can help analysts observe the wider biological system.

Imagine unusual respiratory illness appearing among workers on several farms while veterinary reporting shows animals at the same farms have also been unwell and local media have discussed livestock illness for days, with genomic information later revealing a relationship. If those streams are assessed independently, the significance emerges slowly. A One Health epidemic-intelligence approach asks whether the relationship can become visible sooner — not by asking AI to diagnose zoonotic transmission from news reports, but by ensuring that an animal-health signal is not invisible simply because the analyst is sitting inside a human-health institution. That is the same interface problem we have traced across the law, appearing now in the surveillance layer.

What else is happening: the wider landscape

EIOS is the most consolidated example, but it is part of a fast-moving field, and it is worth seeing the whole picture. In the United States, the CDC Center for Forecasting and Outbreak Analytics, launched in 2022, was explicitly conceived as something like a “National Weather Service for infectious diseases” — using data, models and analytics to inform decisions. In the private sector, BlueDot alerted its clients to the Wuhan cluster on 9 January 2020, roughly nine days before WHO’s public notice, by combining news signals with global airline-travel data to anticipate where cases might appear. Academic and non-profit systems such as HealthMap and ProMED continue to operate, and the research literature on AI-driven early warning is expanding quickly. In Europe, the ECDC runs its own epidemic-intelligence and early-warning functions, and national agencies including the UK’s UKHSA maintain sophisticated surveillance of their own.

The WHO Hub has also moved beyond media scanning: its Collaborative Surveillance framework promotes data and information sharing so that outbreaks can be detected and controlled faster, and in 2023 it launched the International Pathogen Surveillance Network (IPSN) to strengthen genomic surveillance globally. The direction of travel is clear — from isolated tools towards an interconnected intelligence ecosystem. The open question is whether that ecosystem’s governance can keep pace with its sensing.

Algorithms have geography too

There is an important limitation that no amount of computing power removes: open-source intelligence can only analyse information that exists and can be accessed. Some countries have dense digital-media environments; others do not. Some local health events generate substantial online reporting; others occur in communities with limited internet access. Some languages are supported extremely well by automated tools; others are not. Governments differ in transparency, journalistic capacity varies, and internet access can be disrupted deliberately or accidentally. A system monitoring open sources therefore inherits the inequalities of the information environment it observes — a quiet map does not necessarily mean nothing is happening; it may mean nobody is publishing information the system can see.

This is why more data can create more bias rather than less. If one country produces a hundred times more digital content than another, an uncritical system will simply find more signals there; if media attention concentrates on a frightening disease, the sheer volume of reporting can make that threat appear disproportionately important; and if certain communities are underrepresented online, their health problems can remain comparatively invisible. AI does not remove the need to understand surveillance bias. It can amplify it — which is one more reason the human role is a safeguard, not an inconvenience. A signal can be interesting without being true, true without being important, and important without requiring international escalation; those distinctions require evidence, and the verification firewall between machine-assisted detection and authoritative assessment is what makes the whole system trustworthy.

Provenance is the next challenge — and where Solon comes in

This is where EIOS and our own Solon research touch the same wider problem from different directions. EIOS is concerned principally with epidemic intelligence: finding and assessing signals that may indicate emerging threats. Solon is exploring a different question: once a biological event becomes visible, which laws, institutions, obligations and decision thresholds surround it? The systems should not be confused — but they share a defining requirement. An answer needs a route back to its evidence. If AI says an outbreak may be occurring, analysts need to know which reports produced that assessment; if AI says a particular international legal obligation may apply, researchers need to know which treaty provision or official source supports the conclusion. Provenance is not simply a technical feature; it is part of trust. An AI-generated summary can sound authoritative even when the underlying evidence is weak, so every conclusion needs a traceable source and an honest representation of uncertainty — distinguishing confirmed information from plausible but unverified signal, rather than smoothing the difference away with fluent language.

Consider a future surveillance environment in which an EIOS-like system identifies reports of unusual livestock mortality in one country, then days later detects local reports of unexplained illness among agricultural workers; analysts verify both signals; and genomic surveillance subsequently identifies a novel virus. At that point the question changes. Detection asks what is happening? Governance asks who needs to know, which systems should activate, and what happens next? This is where epidemic intelligence and governance intelligence could complement one another — not one giant AI making every decision, but different tools supporting different stages of the same process.

From signal to decision

This connects EIOS directly to the governance-latency framework we have developed. An emerging event travels through a sequence: a biological event, open-source information, machine-assisted detection, analyst assessment, verification, institutional escalation, formal surveillance, and finally a decision. AI may dramatically shorten one interval — but if the remaining transitions are slow, the overall system may improve very little. A faster sensor attached to a slow decision architecture still produces a slow response. This is why health-security technology has to be evaluated as part of a governance system rather than as a gadget: the useful question is not how impressive the interface is, but whether the result reaches the institutions capable of acting on it, quickly enough to matter.

What to watch next

Several developments will decide how much difference AI-assisted epidemic intelligence actually makes. The first is multilingual capability — the earlier a system understands local reporting in the language it first appears in, the greater the potential reduction in detection latency. The second is cross-sector intelligence — human, animal, food and environmental signals need to become easier to examine together. The third is provenance — AI outputs need transparent links to the information supporting them. The fourth is honest uncertainty — systems should distinguish confirmed information from plausible but unverified signal. And the fifth is governance integration — finding a threat faster is useful only if the result reaches the institutions capable of acting on it. Those are measurable questions, and they are harder than demonstrating an impressive AI demo.

Where One Health Security stands

Our position is straightforward, and deliberately independent. The volume of information potentially relevant to emerging biological threats has already exceeded what human analysts can monitor unaided, so technology is becoming part of the surveillance infrastructure whether or not any single institution intends it. That does not mean handing responsibility for outbreak detection to an algorithm; it means using machines to expand what expert humans can see, while protecting the verification, provenance and governance that make the output trustworthy. We think the most valuable work now sits at the interfaces — between epidemic intelligence and governance intelligence, between human, animal and environmental signals, and between a fast sensor and a slower decision architecture. That is the space our Solon research is built to examine, and we would welcome working with the initiatives — national programmes, the WHO Hub, EIOS and the wider community — that are building the sensing side of the same system.

EIOS demonstrates something important about the direction of modern health security: the future of epidemic intelligence is unlikely to be purely human, and it should not be purely artificial either. The interesting system sits between them — the machine watches, and the human decides. And for One Health Security, the next question is already becoming clear: once AI helps us see a biological threat sooner, can our institutions move equally quickly?

Related One Health Security work

This piece connects to our applied-research project Building Solon (governance intelligence for biological risk) and to the Rules of Outbreaks series — in particular The International Health Regulations and The Law Has a One Health Problem. See also What is One Health Security?

Applied Research note, August 2026. Independent analysis by One Health Security — not affiliated with, or endorsed by, WHO, the WHO Hub for Pandemic and Epidemic Intelligence, or the EIOS initiative. Figures (including the approximately 120 countries and 30 organisations using EIOS by the end of 2025) reflect WHO’s public statements at the time of writing.

Questions & Answers

References and further reading

1. World Health Organization. Epidemic Intelligence from Open Sources (EIOS) — initiative, collaboration and technology.
2. World Health Organization. WHO Hub for Pandemic and Epidemic Intelligence (Berlin); and WHO, Collaborative Surveillance.
3. Public Health Agency of Canada. Global Public Health Intelligence Network (GPHIN).
4. International Society for Infectious Diseases, ProMED; and Boston Children’s Hospital, HealthMap.
5. US Centers for Disease Control and Prevention, Center for Forecasting and Outbreak Analytics; and BlueDot.
6. World Health Organization. International Health Regulations (2005), as amended.