The Pulse
Google AI Model Tops CDC’s 2025–26 Flu Forecast Review
Google’s AI-built flu forecasting model ranked first among individual submissions in the CDC’s review of the 2025–26 influenza season. The CDC’s own combined forecast ranked seventh, while its analysis found that forecast accuracy weakened
AI.info Team · Written with AI ·
Google’s AI-built flu forecast ranked first among individual submissions in the CDC’s review of the 2025–26 season, but the agency’s own combined forecast placed seventh. The distinction matters: Google’s result led the field of individual team models, while the CDC’s FluSight ensemble—a separate forecast built by combining submitted models—remained the tool the agency used for its public forecast messaging. The CDC published its season evaluation on September 30, the same day Google announced its result.
The CDC evaluated 39 eligible models, and identified Google_SAI-FluEns as the top-performing individual team submission. Its analysis scored forecasts against observed hospital admissions across U.S. jurisdictions, using relative Weighted Interval Score, or WIS; lower scores indicate better performance. The CDC says 33 of the 39 models beat a baseline that simply carries forward the previous week’s admissions.
Google’s model beat the field, not the CDC ensemble
The CDC’s FluSight ensemble ranked seventh overall, while Google’s model took the top individual-model position. The agency combines forecasts from participating teams for its ensemble, rather than adopting the winning individual model as its own forecast. That makes the result a comparison between two different approaches, not evidence that the CDC replaced its forecast with Google’s.
Forecasts covered the current week and up to three weeks ahead. The CDC’s evaluation excluded national forecasts from its main scoring comparison and excluded Puerto Rico because of data availability; it also omitted models that submitted fewer than 75% of the eligible targets. Google’s September 30 post says the CDC evaluation found its model best matched observed admissions among those 39 eligible models.
ERA generated models; researchers selected the submissions
Google says its forecasts were developed with Empirical Research Assistance, or ERA, a system that uses AI to generate and optimize scientific software. In an earlier description of the work, Google Research said ERA generated candidate forecasting code and that researchers selected models for live weekly submissions. The system was not simply asked to predict the season in one pass: its forecasts were submitted prospectively, before the corresponding hospital-admission figures were known.
“In a fully prospective, real-time evaluation during the 2025–2026 US respiratory season, the system autonomously discovered methodologically diverse models for influenza, COVID-19, and respiratory syncytial virus (RSV).”
Sarah Martinson et al., “Prospective multi-pathogen disease forecasting using autonomous LLM-guided tree search”
The paper describes ERA as an LLM-guided search process that repeatedly generates, tests and refines forecasting code. Researchers provided task instructions, historical surveillance data and an evaluation setup; the system explored candidate approaches, while the team chose models to submit. Google says ERA’s underlying technology is available to trusted testers through its experimental science tools.
Fast changes still challenged forecasts
The CDC classified the 2025–26 flu season as moderate, with weekly hospital admissions peaking nationally during the week ending December 27, 2025, at more than 40,000. Its report says forecast performance declined during periods when influenza trends changed quickly. The CDC’s ensemble prediction intervals failed to anticipate the late-December rise and the mid-January decline.
Google’s research paper also describes limits in its own influenza forecasts: the ensemble tended to underpredict as the season initially progressed and overpredict after the peak. A top overall ranking therefore does not mean the system tracked every turn in the outbreak accurately. The CDC’s evaluation offers a more specific measure of the result: Google’s model led the eligible individual submissions across the full season, while the agency’s combined forecast ranked seventh and struggled around abrupt shifts in admissions.