The flu tracker that stopped checking the flu

In 2008, Google launched Flu Trends. The idea: when more people search for flu symptoms, more people probably have the flu. A paper in the journal Nature in 2009 described how it worked. It picked 45 search terms, out of 50 million candidates, whose ups and downs best matched official flu data from the US Centers for Disease Control and Prevention (CDC).

Early on, it seemed to give a picture of flu activity faster than the CDC, whose reports arrive with a delay of a week or two. But it missed the 2009 swine flu pandemic, which arrived outside the normal flu season, and Google updated the model.

In February 2013, Nature reported that Flu Trends was estimating more than double the share of doctor visits for flu-like illness that the CDC recorded. A 2014 analysis in the journal Science found it had been too high in 100 out of 108 weeks from August 2011, and that a simple forecast based on the CDC's own delayed data did better.

In August 2015, Google stopped publishing Flu Trends estimates and instead gave its data to a few research groups to build their own models.

Flu Trends found search terms that rose and fell with flu season, not terms that were caused by flu. The Science authors noted that some early matches had nothing to do with illness, such as searches about high school basketball, which also peaks in winter. Patterns like that can look strong in past data and fail later.

The world under the model also kept moving. Google itself changed how search worked, for example by suggesting extra search terms, and news coverage of flu can make healthy people search too. The authors called the core mistake "big data hubris": treating a huge pile of data as a replacement for careful measurement, instead of an addition to it.

Today's AI has the same weakness. A model can find patterns that match the past without understanding the cause. Unless someone keeps checking its output against real, independent measurements, it can drift for years while still looking busy and confident.

The idea behind this caseLesson: Check what AI tells you →

1. 1Keep a trusted, independent measure of the truth and compare the AI's numbers against it on a regular schedule.
2. 2Combine AI estimates with the traditional data, rather than replacing the traditional data.
3. 3Ask whether a pattern makes sense as a cause, or only happens to line up with the past.
4. 4Re-check the model whenever the source of its data changes, such as a new app, form or website.
5. 5Publish how the estimate is made, so outsiders can spot problems early.