Google's AI tops the CDC's flu hospital forecast table
Google says a forecasting model built with its AI tools was the most accurate of 39 entries in the CDC's latest FluSight contest to predict US flu hospital admissions.
A flu forecasting model built with Google's AI was the most accurate entry in the US Centers for Disease Control and Prevention's seasonal hospital admissions contest, according to the agency's end-of-season review.
The CDC's FluSight programme collects weekly forecasts of influenza-related hospital admissions from government, industry and academic teams between October and May. Each week, the combined forecast is used to give hospitals and public health officials an early read on the demand heading their way. The 2025-26 analysis, published this week, found that of 39 eligible models, Google's best matched what actually happened.
The tool behind the winning forecasts
Google's entries were produced with Empirical Research Assistance, or ERA, an internal tool that generates optimisation algorithms across scientific fields. Research behind ERA was recently published in the journal Nature, and Google says the underlying technology is now available to trusted testers through its experimental science tools.
Forecast competitions of this kind are one of the cleaner ways to measure AI claims. FluSight scores every model against the same outcome on the same timeline, so the ranking is not a demo or a benchmark the vendor chose for itself — it is a public league table run by a health agency, and it is open to anyone who wants to enter.
The 2025-26 season is unlikely to be the last word. Forecast accuracy varies with the strain of the season, and Google's edge in admissions is narrower in scope than the universal disease forecasting the company gestures at. Still, a first place in a field of 39, judged against real hospital data, is a more useful signal than most AI performance claims.
Why hospital forecasts matter
FluSight exists because winter demand is hard to plan for. If a state's hospitals know a surge is likely three weeks out, they can staff up, move supplies and prepare messaging. The models do not replace surveillance data — they extend it forward, and small improvements in that forward view have direct operational value.
Our opinion
There is a quiet, unglamorous achievement buried in this result. The AI systems that generate the loudest headlines are usually judged on tests their own creators designed; Google's flu model was scored by a federal agency against a season it could not rehearse and a metric — hospital admissions — that cannot be talked up. Coming first out of 39 entries is a real result, not a press release dressed as one.
The honest caveat is that Google is now the vendor and the cheerleader at the same time. ERA's details live behind a Nature paper and a trusted-tester queue, so rivals cannot yet reproduce the win. A league table that everyone can enter is only as meaningful as the code they let everyone see, and until that closes, this should be read as a strong season rather than a settled contest.