“GOOGLE & OTHERS are suppressing voices of Conservatives”, tweeted Donald Trump in 2018. “They are controlling what we can & cannot see.” The president’s charges of bias are often dubious. But many people worry about algorithms absorbing human prejudices. Robert Epstein, an academic, has compiled data that show Google suggesting more positive terms when users type “Hillary Clinton” than when they look up Mr Trump. PJ Media, a conservative blog, claims that liberal sites get 96% of results for “Trump” on Google’s news page, a compilation of links to recent articles.
Google says that the 10,000 human evaluators who rate sources for its search engine assess “expertise” and “trustworthiness” but not ideology. Web-traffic figures support this defence. Sites with high scores from fact-checking groups, whose judgments probably resemble Google’s, draw larger shares of their visitors from search engines than sites with low scores do. Factually inaccurate sources also tend to have strong left- or right-wing slants.
Nonetheless, a subtle bias might not show up in such broad statistics. To test for favouritism, The Economist ran an experiment, comparing a news site’s share of search results with a statistical prediction based on its output, reach and accuracy.
We first wrote a program to obtain Google results for any keyword. Using a browser with no history, in a politically centrist part of Kansas, we searched for 31 terms for each day in 2018, yielding 175,000 links.
Next, we built a model to predict each site’s share of the links Google produces for each keyword, based on the premise that search results should reflect accuracy and audience size, as Google claims. We started with each outlet’s popularity on social media and, using data from Meltwater, a media-tracking firm, how often they covered each topic. We also used accuracy ratings from fact-checking websites, tallies of Pulitzer prizes and results from a poll by YouGov about Americans’ trust in 37 sources.
If Google favoured liberals, left-wing sites would appear more often than our model predicted, and right-wing ones less. We saw no such trend. Overall, centre-left sites like the New York Times got the most links—but only about as many as our model suggested. Fox News beat its modest expectations. Because most far-right outlets had bad trust scores, they got few search results. But so did Daily Kos, a far-left site.
Our study does not prove Google is impartial. In theory, Google could serve un-biased links only to users without a browsing history. If fact-checkers and Pulitzer voters are partisan, our model will be too.
Moreover, some keywords did suggest bias—in both directions. Just as PJ Media charged, the New York Times was over-represented on searches for “Trump”. However, searches for “crime” leaned right: Fox News got far more links than expected.
This implies that Google’s main form of favouritism is to boost viral articles. The most incendiary stories about Mr Trump come from leftist sources. Gory crime coverage is more prevalent on right-leaning sites. Readers will keep clicking on both. ◼