2010 Australian prime minister candidate Julia Gillard

2010 Australian prime minister candidate Julia Gillard

Bidgee/Wikimedia/Creative Commons (CC BY-SA 3.0)

Internet search engines may be influencing elections

“What we’re talking about here is a means of mind control on a massive scale that there is no precedent for in human history.” That may sound hyperbolic, but Robert Epstein says it’s not an exaggeration. Epstein, a research psychologist at the American Institute for Behavioral Research in Vista, California, has found that the higher a politician ranks on a page of Internet search results, the more likely you are to vote for them.

“I have a lot of faith in the methods they’ve used, and I think it’s a very rigorously conducted study,” says Nicholas Diakopoulos, a computer scientist at the University of Maryland, College Park, who was not involved in the research. “I don’t think that they’ve overstated their claims.”

In their first experiment, Epstein and colleagues recruited three groups of 102 volunteers in San Diego, California, who were generally representative of the U.S. voting population in terms of age, race, political affiliation, and other traits. The researchers wanted to know if they could influence who the Californians would have voted for in the 2010 election … for prime minister of Australia.

So they built a fake search engine called Kadoodle that returned a list of 30 websites for the finalist candidates, 15 for Tony Abbott and 15 for Julia Gillard. Most of the Californians knew little about either candidate before the test began, so the experiment was their only real exposure to Australian politics. What they didn’t know was that the search engine had been rigged to display the results in an order biased toward one candidate or the other. For example, in the most extreme scenario, a subject would see 15 webpages with information about Gillard’s platform and objectives followed by 15 similar results for Abbott.

As predicted, subjects spent far more time reading Web pages near the top of the list. But what surprised researchers was the difference those rankings made: Biased search results increased the number of undecided voters choosing the favored candidate by 48% compared with a control group that saw an equal mix of both candidates throughout the list. Very few subjects noticed they were being manipulated, but those who did were actually more likely to vote in line with the biased results. “We expect the search engine to be making wise choices,” Epstein says. “What they’re saying is, ‘Well yes, I see the bias and that’s telling me … the search engine is doing its job.’” 

In a second experiment, the scientists repeated the first test on 2100 participants recruited online through Amazon’s labor crowdsourcing site Mechanical Turk. The subjects were also chosen to be representative of the U.S. voting population. The large sample size—and additional details provided by users—allowed the researchers to pinpoint which demographics were most vulnerable to search engine manipulation: Divorcees, Republicans, and subjects who reported low familiarity with the candidates were among the easiest groups to influence, whereas participants who were better informed, married, or reported an annual household income between $40,000 and $50,000 were harder to sway. Moderate Republicans were the most susceptible of any group: The manipulated search results increased the number of undecided voters who said they would choose the favored candidate by 80%.

“In a two-person race, a candidate can only count on getting half of the uncommitted votes, which is worthless. With the help of biased search rankings, a candidate might be able to get 90% of the uncommitted votes [in select demographics],” Epstein explains.

In a third experiment, the team tested its hypothesis in a real, ongoing election: the 2014 general election in India. After recruiting a sample of 2150 undecided Indian voters, the researchers repeated the original experiment, replacing the Australian candidates with the three Indian politicians who were actually running at the time. The results of the real world trial were slightly less dramatic—an outcome that researchers attribute to voters’ higher familiarity with the candidates. But merely changing which candidate appeared higher in the results still increased the number of undecided Indian voters who would vote for that candidate by 12% or more compared with controls. And once again, awareness of the manipulation enhanced the effect.

A few percentage points here and there may seem meager, but the authors point out that elections are often won by margins smaller than 1%. If 80% of eligible voters have Internet access and 10% of them are undecided, the search engine effect could convince an additional 25% of those undecided to vote for a target candidate, the team reports online this week in the Proceedings of the National Academy of Sciences. That type of swing would determine the election outcome, as long as the expected win margin was 2% or less. “This is a huge effect,” Epstein says. “It’s so big that it’s quite dangerous.”

But perhaps the most concerning aspect of the findings is that a search engine doesn’t even have to intentionally manipulate the order of results for this effect to manifest. Organic search algorithms already in place naturally put one candidate’s name higher on the list than others. This is based on factors like “relevance” and “credibility” (terms that are closely guarded by developers at Google and other major search engines). So the public is already being influenced by the search engine manipulation effect, Epstein says. “Without any intervention by anyone working at Google, it means that Google’s algorithm has been determining the outcome of close elections around the world.”

Presumably Google isn’t intentionally tweaking its algorithms to favor certain presidential candidates, but Epstein says it would extremely difficult to tell if it were. He also points out that the Internet mogul will benefit more from certain election outcomes than others.

And according to Epstein, Google is very aware both of the power it wields, as well as the research his team is doing: When the team recruited volunteers from the Internet in the second experiment, two of the IP addresses came from Google’s head office, he says.

“It’s easy to point the finger at the algorithm because it’s this supposedly inert thing, but there are a lot of people behind the algorithm,” Diakopoulos says. “I think that it does pose a threat to the legitimacy of the democracy that we have. We desperately need to have a public conversation about the role of these systems in the democratic processes.”