AI & Technology
The Model That Changed What It Measured
Google Flu Trends overestimated American flu levels in 100 of 108 consecutive weeks before outside researchers published the tally.
2026-08-022 min read
A model is never evaluated against the world. It is evaluated against whatever record of the world you happened to have — a label, a survey, a lagging report — and that record is what it actually learns to predict. For a while the distinction costs nothing, because the proxy and the truth move together. It starts costing something the moment the model is deployed into the system that produces the proxy, because the thing being measured now has a new input: the measurement.
Google Flu Trends is the cleanest case on record. Launched in 2008 and published in Nature the following year, it inferred influenza-like illness from the volume of flu-related search queries, and it was fitted to reproduce the CDC's surveillance reports — not infection rates, which nobody observes directly. It was widely praised, it missed the 2009 H1N1 outbreak, and then it drifted: Lazer, Kennedy, King and Vespignani reported in Science in 2014 that it overshot the CDC figure in 100 of 108 consecutive weeks beginning in August 2011, at the 2012–13 peak by roughly double. Their diagnosis had two parts. The first they named big-data hubris — treating query volume as a substitute for a model of the disease rather than a supplement to one. The second is the more uncomfortable, because it implicates no analyst: Google kept changing its own search product, adding suggestions and related-search prompts that led people to search for flu terms they would not otherwise have typed. The sensor and the product were the same box. Google stopped publishing the estimates in 2015.
That second failure has no fix inside the model. Retraining on more data cannot help when the data-generating process is reacting to the deployment, and accuracy computed against the proxy looks fine right up until it doesn't. What the case argues for is knowing which of your numbers is ground truth and which is a shadow of it, then checking on a schedule whether the shadow has started moving on its own.
Key insights
A model learns to predict your record of the world, not the world — and the two agree only until something moves them apart.
Google Flu Trends overshot CDC influenza figures in 100 of 108 consecutive weeks, at the 2012–13 peak by roughly double.
When a deployed model changes the behavior that produces its own training signal, more data cannot fix it.
Why it matters
Every model in production is scored against a proxy, and the proxy is usually the thing your team controls least and inspects least.
Once a model influences the behavior that generates its own training signal, its error stops being random and starts compounding in one direction.
Use this tomorrow
1Write down, for the one model your team depends on most, the exact field it was trained to predict and the real-world outcome you actually care about; if those come out as the same sentence, you have not found the proxy yet.
2Pull the last six months of that model's accuracy numbers and mark which were computed against the proxy rather than an independently collected sample; if none were independent, book one independent measurement before the quarter closes.
Go deeper
The textbook name for the second failure is feedback, but the practical version is more specific: the deployment and the measurement share a channel. Google's search box was simultaneously the sensor collecting flu signal and a product being actively optimised for engagement, and changes made for the second reason quietly corrupted the first. Any recommender, pricing model, or fraud scorer that acts on the same population it measures has this structure. The mitigation is boring and effective — hold out a slice of traffic the model never touches, so at least one measurement stays uncontaminated.
Google Flu Trends is usually retold as a story about big data failing, which is the wrong lesson and part of why it keeps recurring. Lazer and his co-authors were explicit that the search data was genuinely informative: a model combining query volume with lagged CDC data outperformed either input alone. The failure was the decision to treat a correlational proxy as a replacement for surveillance rather than an early-warning supplement to it. Systems that pair a fast noisy signal with a slow reliable one tend to survive; systems that retire the slow one do not.
