Google has found another job for artificial intelligence, and this one could have real consequences for public health. An AI-powered flu forecasting model from the company finished first in a CDC evaluation of systems predicting influenza hospitalizations across the United States.
The result comes directly from the CDC’s evaluation of the 2025-2026 FluSight season. Researchers submitted weekly forecasts for flu-related hospital admissions, covering the current week and as many as three weeks into the future. The forecasts covered states and Washington, D.C., giving health officials an idea of where hospital demand could be heading.
A total of 34 teams contributed 53 models during the season, although only 39 met the CDC’s requirements for inclusion in its final analysis. Among those models, Google_SAI-FluEns came out on top.
That is particularly interesting because the CDC has its own FluSight ensemble, which combines forecasts from participating models and is used by the agency when communicating its flu forecasts. That ensemble ranked seventh overall. Of the 39 models evaluated, 33 performed better than the CDC’s basic baseline model.
Google explains that its flu forecasts were developed using AI-assisted scientific research. The company used a system called Empirical Research Assistance, or ERA, which can generate optimization algorithms for scientific problems. Google says the technology behind ERA is now being made available to trusted testers through its experimental science tools.
This isn’t simply a case of Google declaring its own AI the winner, either. The ranking comes from the CDC’s evaluation, which compared forecasts against the hospital admissions that actually occurred during the season.
There are some important limits. Flu forecasting remains difficult, particularly when infection patterns suddenly change. The CDC found that forecast performance declined during periods of rapidly changing flu activity. Its FluSight ensemble failed to anticipate some of the sharp increase in hospitalizations around the late-December peak and the decline that followed in January.
The CDC evaluated models primarily using a metric called relative weighted interval score, or WIS. In simple terms, it considers how well a forecast’s predicted range lines up with what eventually happens, with lower scores representing better performance. Google had the best overall result among the individual team submissions.
This doesn’t mean Google AI can tell you whether you’ll get the flu next Tuesday, nor does it make traditional epidemiology obsolete. These are population-level forecasts designed to help anticipate hospital demand, and even the best-performing system can struggle when a flu season suddenly changes direction.
Still, finishing first in a CDC-run evaluation is a lot more convincing than a carefully selected company benchmark. If AI systems can consistently improve forecasts of where disease-related hospitalizations are heading, hospitals and public health agencies could get more time to prepare for surges.
For all the attention surrounding AI-generated images, chatbots, and agents, predicting how many people may soon need a hospital bed feels like a considerably better test of what this technology can actually do.
Support independent tech journalism
NERDS.xyz is independently owned and operated. If you enjoy my coverage of Linux, AI, hardware, cybersecurity, and tech culture, consider supporting the site on Ko-fi.
Support NERDS.xyz
