Google's Flu Forecasting Model Ranks First in CDC Evaluation

Google said this week that a flu forecasting model it built outperformed every other entry judged in the Centers for Disease Control and Prevention’s FluSight evaluation for the 2025-26 flu season. The CDC runs FluSight from October through May each year, combining weekly submissions from government, industry and academic teams that predict United States hospital admissions for the current week and three weeks ahead. The agency uses the combined forecast to tell states and hospitals what demand for care to expect.

The end of season analysis, published this week, measured 39 eligible models against the season’s actual recorded hospital admissions. Google’s submission matched those figures more closely than any of the others. The company built its forecast using a tool it calls Empirical Research Assistance, described as software that generates optimization algorithms across scientific fields rather than a single purpose epidemiology model. Google said the underlying technology behind that tool was recently described in a paper published in the journal Nature, and that it is now available to a limited group of outside testers through its experimental science program.

Google framed the result as evidence that combining AI systems with researchers’ judgment can improve disease forecasting generally, a claim the company has also made about other science applications it has pursued this year, including work on materials design and space infrastructure.

The numbers

The CDC’s FluSight programme drew submissions from 39 eligible models for the 2025-26 season, spanning government, industry and academic teams. Google’s entry, built on its Empirical Research Assistance tool, was the one the CDC’s end of season review found closest to the season’s actual hospital admission counts. The evaluation ran across the full October to May flu season, scoring each submission on both the current week’s predicted admissions and a three week forward outlook, which is the window hospitals use to plan staffing.

No pricing or availability details accompanied the announcement. Google has not said whether the forecasting model itself, as opposed to the underlying research tool, will be offered to health agencies outside the CDC pilot, or whether future seasons will see it submit again under the same evaluation rules. For now the claim rests on one season’s results judged by a single federal programme, not on a published independent audit of the model’s code or training data.