Bwindi Impenetrable National Park in southwestern Uganda is famous for the endangered mountain gorillas that inhabit its lush, misty highlands. It’s also home to around 120 other species of mammal, 350 types of bird and hundreds of plant varieties — and, on its borders, people and livestock. That’s a recipe for what epidemiologists call spillover, the transfer of pathogens between species.

In 2003, the non-profit organization Conservation Through Public Health (CTPH) in Entebbe, Uganda, began work to protect the gorillas (Gorilla beringei beringei) from diseases, such as scabies, that were traced to their human neighbours. The team uses a holistic One Health approach that balances the needs of people with the health of the animals and the ecosystems that surround them. Typically, the process has been reactive: responding to outbreaks as they emerge, says Ssali Ronald Ogwal, a public-health specialist at the CTPH. But the organization is now shifting towards a more proactive, disease-prevention approach — thanks, in part, to artificial intelligence.

COVID’s origins: what we do and don’t know

As part of a three-year collaboration that ended in March 2026 with public-health initiative NESTLER, a joint project between the European Union and African nations, Ogwal collected samples from cattle and poultry around Bwindi. He sent the specimens to a laboratory in Entebbe to be tested for diseases such as brucellosis and Rift Valley fever, both of which can infect people. After sharing the results with the relevant local communities, the team combined the data with the CTPH’s routine gorilla-monitoring records and passed the information on to colleagues at NESTLER. There, the data are being used to train predictive AI models, which are designed to be early-warning systems for disease outbreaks that can affect people, livestock and gorillas.

Many infectious diseases that affect humans are zoonotic — that is, they originate in other animals. This includes such scourges as SARS-CoV-2, the virus that caused the COVID-19 pandemic, and ebolaviruses, which are thought to have originated in fruit bats of the Pteropodidae family. Humans can catch ebolaviruses directly from bats’ bodily fluids, or by way of animals that have come into contact with infected bats.

Agriculture, climate shifts and land-use changes such as deforestation have increased contact between people and wildlife, making spillover events more likely to occur. At the same time, globalization and geopolitics enable diseases to be spread more quickly between humans. As the COVID-19 pandemic so powerfully demonstrated, spillover events can have profound impacts, affecting physical and mental health, livelihoods and education around the world.

For researchers such as Ogwal, AI — in combination with practical measures such as regulating wildlife trading — offers a powerful technology for managing zoonotic diseases and reducing the risk of spillover, possibly even preventing outbreaks in the first place. At least, that’s the theory: AI tools are still limited and are not yet widely used in infectious-disease epidemiology. But the continuing boom in machine-learning techniques is beginning to change that.

“Rather than just waiting for outbreaks to happen, these AI machines support a deep analysis of large volumes of data to identify patterns,” says Ogwal. Predictions can be acted on “before anything escalates”.

AI virus sleuths

If you want to staunch outbreaks before they get out of hand, quick and efficient processing of data is key.

AI tools such as machine-learning models are designed for deftly “sorting through massive amounts of data, following rules, finding patterns”, says Edward Holmes, a virologist at the University of Sydney in Australia, who investigates the metagenomics of environmental samples.

Stopping the next flu pandemic

Holmes develops machine-learning models to identify new zoonotic viruses in samples that were not necessarily collected for that purpose. He uses sequencing data from public resources such as GenBank or Pathoplexus, an open-source database of viral pathogens. The models are trained on the sequences of all known human pathogens, so they can recognize traits associated with disease emergence in people. “The cell receptors, modes of transmission — all those sorts of things play into these AI algorithms,” Holmes explains. The models then learn to identify what Holmes terms “risky” viruses that researchers can look out for in the field, and even design prophylactic vaccines against them.

One such model is LucaProt, a deep-learning algorithm designed for RNA virus discovery using protein structures derived from sequence data, developed by Holmes and collaborators. In 2024, the researchers applied LucaProt to a collection of 10,487 public metatranscriptome data sets, identifying 161,979 species of RNA virus. Among those were 70,458 viruses that had never been seen before — the largest discovery of viruses ever recorded in a single study, according to the authors¹. Although few, if any, are likely to be harmful to humans, researchers say that understanding emerging infectious diseases requires comprehending the diversity of viruses and how they jump between species and evolve.

Other researchers are also applying AI to virus discovery. In 2021, scientists at the University of Glasgow, UK, described a machine-learning method that used known zoonotic viruses to predict which new viruses could potentially cross over into humans². And earlier this year, researchers in the United States described a predictive model that identifies disease hosts and suggests when sampling would be most productive³.

Such models are being used mostly for research rather than to inform surveillance, but Holmes suggests that regular sampling for specific antibodies could be set up along “fault lines” — places where spillover is likely to happen, such as live animal markets, poultry farms or settlements near bat roosts. “I’d have local staff trained to generate and analyse the data,” Holmes says. “And that would all go to a central location where you’d have a global radar map of what’s going on, like air traffic control.” One such centre might be the World Health Organization’s Hub for Pandemic and Epidemic Intelligence in Berlin, he suggests.

This type of set-up is feasible, Holmes says, and cost-effective. In 2022, the World Bank estimated that preventions guided by the One Health approach would cost up to US$11.5 billion a year — about one-third the cost of managing pandemics⁴. “It’s a politics and people problem,” Holmes says, noting that the short-term thinking of modern politics often prevails over the potential long-term financial savings of such measures.

And humans would still have to be in the loop for surveillance, Holmes and others say. “Combining AI with metagenomic sequencing is valuable but cannot, by itself, resolve the fundamental uncertainties in pathogen emergence,” wrote Nader Ebrahimi and Amir Ghaemi, virologists at the Pasteur Institute of Iran in Tehran, in The Lancet Infectious Diseases in January⁵.

Disease surveillance

Beyond virus identification, AI can also be used to forecast and monitor outbreaks. BlueDot, a firm in Toronto, Canada, that assesses infectious-disease risk, provides these services for clients such as the Gulf Center for Disease Prevention and Control in Riyadh and the City of Chicago, Illinois. The company uses AI to collect and filter thousands of articles and official data from public-health organizations, translating 65 languages to pull out information on specific diseases and symptoms. It then integrates further sources such as air-travel ticket sales to advise clients that “these are the risks that are most connected to your location, based on the way that people are moving”, explains Andrea Thomas, the company’s vice-president of epidemiology and data science.

In 2015, researchers working for BlueDot identified Miami, Florida, as one of several possible Zika virus outbreak locations across North and South America by combining ecological data for the Aedes mosquitoes that transmit Zika, temperature profiles and information on flights from an outbreak site in Brazil⁶. Their prediction proved prescient over the following few years, when 1,471 cases of the disease were documented in Florida⁷.

Monitoring fruit bat colonies could provide early warning for Nipah outbreaks

Another BlueDot project built a boosted regression tree model — a form of AI that combines predictions from several models — to identify regions at future risk of dengue fever, chikungunya, Zika and other mosquito-borne illnesses. The team used data on mosquito species and their habitats, alongside climate-change predictions, to assess how mosquito distribution might change over time. Looking ahead to 2036, they have identified probable patterns of disease risk across Europe, the United States and southern Canada (see go.nature.com/4xpyryl).

Although it has grown more sophisticated over the past decade, AI use in global health monitoring isn’t new. Twenty years ago, epidemiologist John Brownstein and his colleagues at Boston Children’s Hospital in Massachusetts created a Google Maps-based tool called HealthMap to monitor disease outbreaks around the world (see go.nature.com/4wlarvg). The system uses specialized data dictionaries — describing, for instance, what diseases are called in different countries — to collate the data. But the team is now transitioning to an approach based on large language models (LLMs), which have drastically changed how HealthMap collates data, Brownstein says.

HealthMap pulls data from sources such as news websites, government health departments and social media, and uses Fisher–Robinson Bayesian filtering — a form of AI used for identifying spam e-mails — to weed out noise. In late 2019, the system sent out the first alert anywhere in the world relating to the disease later identified as COVID-19. “Informal sources of infectious-disease surveillance could provide this early window into what is happening in a population,” Brownstein explains.

Since 2025, the team has branched out into other projects, including BEACON (Biothreats Emergence, Analysis and Communications Network), based at the Center on Emerging Infectious Diseases at Boston University. BEACON combines alerts from HealthMap and other sources using an LLM to create a dashboard that helps to refine advice and information about diseases such as the current Ebola outbreak in the Democratic Republic of the Congo and Uganda. “It’s a combination of both AI and human experts,” says Brownstein. “So it’s a much more in-depth, contextualized assessment of risk.” More than 227,000 users from 233 countries and territories, including public-health professionals and clinicians, have used the site since its launch, Brownstein says.

At BlueDot, LLMs have simplified the way that the company shares and tailors information for its clients, Thomas says. “All of the information is just becoming easier to gather and distil using the newer technologies.” And, she adds, the near-democratization of modelling makes it easier for non-experts to layer in their own data without having to share the information externally.

Good use

But, as with AI use in general, there are also ethical issues to consider.

One concern about AI use in detecting zoonotic diseases is how it includes — or doesn’t include — local communities, particularly those in low- and middle-income countries. Researchers, including epidemiologists and socio-anthropologists, worry that AI bias could occur if community members’ lived experiences, local knowledge and realities on the ground are not incorporated into the models.