The pattern underneath
In 2008–2009 Google published, including in Nature, a model trained on tens of millions of queries. A handful of search terms tracked CDC flu reports and seemed to see the wave a week or two early. Then the engine changed (suggestions, related searches), journalists wrote about flu, people searched because of the news rather than the virus, and the model — overfit to old winters — drifted. In 2012–2013 it overshot official rates badly. Lazer and colleagues called it ‘big data hubris’ in Science (2014): volume is not understanding, and a platform that edits itself is not a stable instrument. The public Flu Trends service ended in 2015.
How the phrase gets misused
Used as proof that ‘search knows the population better than doctors’ — or as proof that all digital epidemiology is a joke. Both are lazy. The method is a sensor. Sensors drift. They need a slower, duller instrument (surveys, clinics) running beside them.
A more precise frame
For NOOS: treat query lines as weather. Treat the result page as the room the person walks into. Never publish a national ‘sickness score’ from search alone. Keep Flu Trends on the desk as a warning, not a mascot.