Google's AI just became the best flu forecaster the CDC has ever worked with

Google's AI model ranked #1 among 39 eligible submissions in the CDC's FluSight program for the 2025-26 flu season, most accurately matching the season's actual observed hospital admissions compared to every other participating forecast.

Why it matters: FluSight isn't a casual leaderboard. It's a real forecasting competition that feeds directly into how the CDC communicates expected flu demand to hospitals and health systems across the country, so ranking first means Google's model genuinely outperformed dozens of other government, industry, and academic forecasting teams at a task with real practical stakes.

The model forecasts flu-related hospital admissions in the United States, covering both the current week and up to three weeks ahead, broken down at the state level. That forecast then becomes part of a combined, ensemble forecast the CDC uses to estimate anticipated demand for medical services across different regions.

In real life it's the difference between a hospital finding out it's overwhelmed with flu patients the day it happens, versus having a few weeks' warning to actually staff up and prepare beds in advance.

Google built the model using Empirical Research Assistance (ERA), a new AI tool designed to generate optimization algorithms across different scientific fields, not just public health specifically. The underlying technology behind ERA was recently published in the journal Nature, giving it real peer-reviewed scientific backing rather than just an internal Google claim.

• Tool: Empirical Research Assistance (ERA).
• Purpose: generates optimization algorithms across scientific domains.
• Validation: underlying technology published in Nature.
• Access: available to trusted testers via labs.google/science.

Why it matters: ERA being a general-purpose scientific optimization tool, not something built narrowly just for flu forecasting, suggests this win is a demonstration of a broader capability rather than a one-off specialized project.

Background: FluSight has run since the 2013-14 flu season, making it more than a decade-old effort by the CDC to evaluate and combine flu forecasting models from participating teams. Each season, from October through May, government, industry, and academic teams submit weekly forecasts that the CDC combines into an ensemble prediction.

Why it matters: a program running this long, with this many competing teams each year, provides a genuinely rigorous benchmark. A model claiming the top spot in a field with over a decade of accumulated competition and refinement is a meaningfully stronger claim than winning a brand-new, less battle-tested evaluation.

Who's affected: hospitals and public health officials are the direct users of these forecasts, not patients or the general public. More accurate hospital admission predictions give health systems more reliable advance notice to adjust staffing levels and prepare bed capacity ahead of an actual surge in flu patients, rather than reacting once admissions have already climbed.

The forecast itself doesn't control outcomes, it's the officials and hospital administrators using it who make the actual staffing and resource decisions. But a more accurate input naturally makes those downstream decisions more likely to match real demand.

A Google Research post framed the result as validating a bigger idea: "This performance validates our confidence that AI combined with human ingenuity will improve our ability to forecast diseases worldwide." That's a notably broader claim than just "we're good at flu forecasting specifically."

Why it matters: if Google genuinely intends to apply the same ERA-based approach to forecasting other diseases beyond flu, this result functions as a proof of concept for a much larger ambition, improving disease forecasting broadly, rather than a standalone achievement limited to one seasonal illness.

By the numbers: the CDC estimates that flu results in somewhere between 9.3 million and 41 million illnesses, 100,000 to 710,000 hospitalizations, and 4,900 to 51,000 deaths in the US annually, depending on the severity of a given season.

The most recent finalized full season, 2023-24, was associated with roughly 40 million symptomatic illnesses, 18 million medical visits, 470,000 hospitalizations, and 28,000 deaths.

That scale carries a real economic cost too. Influenza among adults alone was linked to an estimated $29 billion in total economic burden during the 2023-24 season, split between roughly $16 billion in direct healthcare costs and $13 billion in lost productivity.

Why it matters: those wide ranges reflect how dramatically flu severity can swing from one season to the next, which is exactly why accurate forecasting matters so much.

A hospital system that can reliably anticipate whether a given season is trending toward the lower or higher end of that range, a few weeks ahead of time, can make meaningfully better staffing and resource decisions than one relying on guesswork alone.

Background: yes, flu forecasting joins a growing list of Google AI projects aimed at scientific and public health problems rather than consumer products. Google's AlphaFold system, which predicts protein structures and won a share of the 2024 Nobel Prize in Chemistry, is now used by more than 4 million researchers across 190 countries for work ranging from drug discovery to studying neglected diseases.

Google has also pushed AI into weather forecasting, releasing a model in September 2026 capable of hourly forecasts instead of the standard six-hour intervals, with flood forecasting now covering more than 2 billion people across 150 countries. A separate AI co-scientist system, announced in February 2025, is designed to help generate and test scientific hypotheses directly.

Why it matters: flu forecasting fits neatly into that same pattern: applying frontier AI research to genuinely consequential scientific and public health problems, then publishing peer-reviewed validation of the results rather than just marketing claims. It's a meaningfully different track than Google's consumer AI products, even when both draw on related underlying technology.

• You won't interact with this model directly, but it's part of the infrastructure that helps hospitals near you prepare for flu season demand more accurately.
• This is a genuine example of AI being applied to unglamorous, practical public health infrastructure, not a flashy consumer product, worth noting amid the constant stream of AI announcements focused on chatbots and image generation.
• If you work in public health, epidemiology, or research more broadly, ERA's limited tester access might be worth tracking if Google expands availability, given its Nature-published, peer-reviewed foundation.
• Watch for Google extending this approach to other diseases. Google's own framing suggests flu forecasting may be the first of several disease-forecasting applications for this technology, not the only one.