America’s next president could be eased into office not just by TV ads or speeches, but by Google’s secret decisions, and no one—except for me and perhaps a few other obscure researchers—would know how this was accomplished.
Research I have been directing in recent years suggests that Google, Inc., has amassed far more power to control elections—indeed, to control a wide variety of opinions and beliefs—than any company in history has ever had. Google’s search algorithm can easily shift the voting preferences of undecided voters by 20 percent or more—up to 80 percent in some demographic groups—with virtually no one knowing they are being manipulated, according to experiments I conducted recently with Ronald E. Robertson.
Given that many elections are won by small margins, this gives Google the power, right now, to flip upwards of 25 percent of the national elections worldwide. In the United States, half of our presidential elections have been won by margins under 7.6 percent, and the 2012 election was won by a margin of only 3.9 percent—well within Google’s control.
There are at least three very real scenarios whereby Google—perhaps even without its leaders’ knowledge—could shape or even decide the election next year. Whether or not Google executives see it this way, the employees who constantly adjust the search giant’s algorithms are manipulating people every minute of every day. The adjustments they make increasingly influence our thinking—including, it turns out, our voting preferences.
What we call in our research the Search Engine Manipulation Effect (SEME) turns out to be one of the largest behavioral effects ever discovered. Our comprehensive new study, just published in the Proceedings of the National Academy of Sciences (PNAS), includes the results of five experiments we conducted with more than 4,500 participants in two countries. Because SEME is virtually invisible as a form of social influence, because the effect is so large and because there are currently no specific regulations anywhere in the world that would prevent Google from using and abusing this technique, we believe SEME is a serious threat to the democratic system of government.
According to Google Trends, at this writing Donald Trump is currently trouncing all other candidates in search activity in 47 of 50 states. Could this activity push him higher in search rankings, and could higher rankings in turn bring him more support? Most definitely—depending, that is, on how Google employees choose to adjust numeric weightings in the search algorithm. Google acknowledges adjusting the algorithm 600 times a year, but the process is secret, so what effect Mr. Trump’s success will have on how he shows up in Google searches is presumably out of his hands.
Our new research leaves little doubt about whether Google has the ability to control voters. In laboratory and online experiments conducted in the United States, we were able to boost the proportion of people who favored any candidate by between 37 and 63 percent after just one search session. The impact of viewing biased rankings repeatedly over a period of weeks or months would undoubtedly be larger.
In our basic experiment, participants were randomly assigned to one of three groups in which search rankings favored either Candidate A, Candidate B or neither candidate. Participants were given brief descriptions of each candidate and then asked how much they liked and trusted each candidate and whom they would vote for. Then they were allowed up to 15 minutes to conduct online research on the candidates using a Google-like search engine we created called Kadoodle.
Each group had access to the same 30 search results—all real search results linking to real web pages from a past election. Only the ordering of the results differed in the three groups. People could click freely on any result or shift between any of five different results pages, just as one can on Google’s search engine.
When our participants were done searching, we asked them those questions again, and, voilà: On all measures, opinions shifted in the direction of the candidate who was favored in the rankings. Trust, liking and voting preferences all shifted predictably.
More alarmingly, we also demonstrated this shift with real voters during an actual electoral campaign—in an experiment conducted with more than 2,000 eligible, undecided voters throughout India during the 2014 Lok Sabha election there—the largest democratic election in history, with more than 800 million eligible voters and 480 million votes ultimately cast. Even here, with real voters who were highly familiar with the candidates and who were being bombarded with campaign rhetoric every day, we showed that search rankings could boost the proportion of people favoring any candidate by more than 20 percent—more than 60 percent in some demographic groups.
Given how powerful this effect is, it’s possible that Google decided the winner of the Indian election. Google’s own daily data on election-related search activity (subsequently removed from the Internet, but not before my colleagues and I downloaded the pages) showed that Narendra Modi, the ultimate winner, outscored his rivals in search activity by more than 25 percent for sixty-one consecutive days before the final votes were cast. That high volume of search activity could easily have been generated by higher search rankings for Modi.
Google’s official comment on SEME research is always the same: “Providing relevant answers has been the cornerstone of Google’s approach to search from the very beginning. It would undermine the people’s trust in our results and company if we were to change course.”
Could any comment be more meaningless? How does providing “relevant answers” to election-related questions rule out the possibility of favoring one candidate over another in search rankings? Google’s statement seems far short of a blanket denial that it ever puts its finger on the scales.
There are three credible scenarios under which Google could easily be flipping elections worldwide as you read this:
First, there is the Western Union Scenario: Google’s executives decide which candidate is best for us—and for the company, of course—and they fiddle with search rankings accordingly. There is precedent in the United States for this kind of backroom king-making. Rutherford B. Hayes, the 19th president of the United States, was put into office in part because of strong support by Western Union. In the late 1800s, Western Union had a monopoly on communications in America, and just before to the election of 1876, the company did its best to assure that only positive news stories about Hayes appeared in newspapers nationwide. It also shared all the telegrams sent by his opponent’s campaign staff with Hayes’s staff. Perhaps the most effective way to wield political influence in today’s high-tech world is to donate money to a candidate and then to use technology to make sure he or she wins. The technology guarantees the win, and the donation guarantees allegiance, which Google has certainly tapped in recent years with the Obama administration.
Given Google’s strong ties to Democrats, there is reason to suspect that if Google or its employees intervene to favor their candidates, it will be to adjust the search algorithm to favor Hillary Clinton. In 2012, Google and its top executives donated more than $800,000 to Obama but only $37,000 to Romney. At least six top tech officials in the Obama administration, including Megan Smith, the country’s chief technology officer, are former Google employees. According to a recent report by the Wall Street Journal, since Obama took office, Google representatives have visited the White House ten times as frequently as representatives from comparable companies—once a week, on average.
Hillary Clinton clearly has Google’s support and is well aware of Google’s value in elections. In April of this year, she hired a top Google executive, Stephanie Hannon, to serve as her chief technology officer. I don’t have any reason to suspect Hannon would use her old connections to aid her candidate, but the fact that she—or any other individual with sufficient clout at Google—has the power to decide elections threatens to undermine the legitimacy of our electoral system, particularly in close elections.
This is, in any case, the most implausible scenario. What company would risk the public outrage and corporate punishment that would follow from being caught manipulating an election?
Second, there is the Marius Milner Scenario: A rogue employee at Google who has sufficient password authority or hacking skills makes a few tweaks in the rankings (perhaps after receiving a text message from some old friend who now works on a campaign), and the deed is done. In 2010, when Google got caught sweeping up personal information from unprotected Wi-Fi networks in more than 30 countries using its Street View vehicles, the entire operation was blamed on one Google employee: software engineer Marius Milner. So they fired him, right? Nope. He’s still there, and on LinkedIn he currently identifies his profession as “hacker.” If, somehow, you have gotten the impression that at least a few of Google’s 37,000 employees are every bit as smart as Milner and possess a certain mischievousness—well, you are probably right, which is why the rogue employee scenario isn’t as far-fetched as it might seem.
And third—and this is the scariest possibility—there is the Algorithm Scenario: Under this scenario, all of Google’s employees are innocent little lambs, but the software is evil. Google’s search algorithm is pushing one candidate to the top of rankings because of what the company coyly dismisses as “organic” search activity by users; it’s harmless, you see, because it’s all natural. Under this scenario, a computer program is picking our elected officials.
To put this another way, our research suggests that no matter how innocent or disinterested Google’s employees may be, Google’s search algorithm, propelled by user activity, has been determining the outcomes of close elections worldwide for years, with increasing impact every year because of increasing Internet penetration.
SEME is powerful precisely because Google is so good at what it does; its search results are generally superb. Having learned that fact over time, we have come to trust those results to a high degree. We have also learned that higher rankings mean better material, which is why 50 percent of our clicks go to the first two items, with more than 90 percent of all clicks going to that precious first search page. Unfortunately, when it comes to elections, that extreme trust we have developed makes us vulnerable to manipulation.
In the final days of a campaign, fortunes are spent on media blitzes directed at a handful of counties where swing voters will determine the winners in the all-important swing states. What a waste of resources! The right person at Google could influence those key voters more than any stump speech could; there is no cheaper, more efficient or subtler way to turn swing voters than SEME. SEME also has one eerie advantage over billboards: when people are unaware of a source of influence, they believe they weren’t being influenced at all; they believe they made up their own minds.
Republicans, take note: A manipulation on Hillary Clinton’s behalf would be particularly easy for Google to carry out, because of all the demographic groups we have looked at so far, no group has been more vulnerable to SEME—in other words, so blindly trusting of search rankings—than moderate Republicans. In a national experiment we conducted in the United States, we were able to shift a whopping 80 percent of moderate Republicans in any direction we chose just by varying search rankings.
There are many ways to influence voters—more ways than ever these days, thanks to cable television, mobile devices and the Internet. Why be so afraid of Google’s search engine? If rankings are so influential, won’t all the candidates be using the latest SEO techniques to make sure they rank high?
SEO is competitive, as are billboards and TV commercials. No problem there. The problem is that for all practical purposes, there is just one search engine. More than 75 percent of online search in the United States is conducted on Google, and in most other countries that proportion is 90 percent. That means that if Google’s CEO, a rogue employee or even just the search algorithm itself favors one candidate, there is no way to counteract that influence. It would be as if Fox News were the only television channel in the country. As Internet penetration grows and more people get their information about candidates online, SEME will become an increasingly powerful form of influence, which means that the programmers and executives who control search engines will also become more powerful.
Worse still, our research shows that even when people do notice they are seeing biased search rankings, their voting preferences still shift in the desired directions—even more than the preferences of people who are oblivious to the bias. In our national study in the United States, 36 percent of people who were unaware of the rankings bias shifted toward the candidate we chose for them, but 45 percent of those who were aware of the bias also shifted. It’s as if the bias was serving as a form of social proof; the search engine clearly prefers one candidate, so that candidate must be the best. (Search results are supposed to be biased, after all; they’re supposed to show us what’s best, second best, and so on.)
Biased rankings are hard for individuals to detect, but what about regulators or election watchdogs? Unfortunately, SEME is easy to hide. The best way to wield this type of influence is to do what Google is becoming better at doing every day: send out customized search results. If search results favoring one candidate were sent only to vulnerable individuals, regulators and watchdogs would be especially hard pressed to find them.
For the record, by the way, our experiments meet the gold standards of research in the behavioral sciences: They are randomized (which means people are randomly assigned to different groups), controlled (which means they include groups in which interventions are either present or absent), counterbalanced (which means critical details, such as names, are presented to half the participants in one order and to half in the opposite order) and double-blind (which means that neither the subjects nor anyone who interacts with them has any idea what the hypotheses are or what groups people are assigned to). Our subject pools are diverse, matched as closely as possible to characteristics of a country’s electorate. Finally, our recent report in PNAS included four replications; in other words, we showed repeatedly—under different conditions and with different groups—that SEME is real.
Our newest research on SEME, conducted with nearly 4,000 people just before the national elections in the UK this past spring, is looking at ways we might be able to protect people from the manipulation. We found the monster; now we’re trying to figure out how to kill it. What we have learned so far is that the only way to protect people from biased search rankings is to break the trust Google has worked so hard to build. When we deliberately mix rankings up, or when we display various kinds of alerts that identify bias, we can suppress SEME to some extent.
It’s hard to imagine Google ever degrading its product and undermining its credibility in such ways, however. To protect the free and fair election, that might leave only one option, as unpalatable at it might seem: government regulation.