As the 2020 presidential election heats up, alarmism over “Big Tech”‘s influence on voters is again reaching a fever pitch. Democrats fear the impact of fake news and Russian bots, while Republicans complain of supposed anti-conservative bias and censorship at Google, Twitter, and Facebook. They all agree that the federal government needs more power over online speech.
To bolster allegations of liberal activism by tech companies, Republicans have put a spotlight on the research of Robert Epstein, a Harvard-trained psychologist and self-described Hillary Clinton supporter who has called Google’s search engine “the most powerful mind control device ever devised.”
A former editor-in-chief of Psychology Today, Epstein has conducted a series of studies over the past seven years that he says show that search engines and social media companies have used subliminal manipulation to shift a significant number of votes to Democrats. He says that the tech industry poses an even more pervasive threat in 2020 to the presidential election and to American democracy.
Google has called his work “nothing more than a poorly constructed conspiracy theory,” while Donald Trump cited Epstein’s research as an explanation for why he lost the popular vote in 2016 by nearly 3 million votes. Hillary Clinton responded that Epstein’s studies have been “debunked.”
A careful review of Epstein’s work reveals that, though his research does not display any obvious methodological or statistical flaws, his findings don’t support the claims that he’s been making about the tech giants’ influence on real elections. Epstein over-extrapolates from his research to make allegations that have been seized on by politicians when making the case for more government control over internet platforms.
In testimony last year before the Senate Judiciary Committee that Sen. Ted Cruz (R–Texas) called “incredibly powerful and incredibly concerning,” Epstein asserted that in 2016, bias in Google’s search algorithm may have shifted between 2.6 and 10.4 million votes to Hillary Clinton; that in 2018, the search engine may have shifted upwards of 78.2 million votes to Democratic candidates; and that if all the tech platforms get together in 2020, they could shift 15 million votes to Joe Biden, engineering a 20 percent win margin and a landslide election victory.
“Election by election, I can calculate how many votes these companies can shift,” Epstein tells Reason. His numeric estimates, he says, come from combining the results of several research studies that date back to 2013.
In Epstein’s initial lab-based and online experiments, participants spent up to 15 minutes researching details on a foreign election they knew nothing about. They used a fake search engine similar to Google that reordered results to favor a particular candidate. Then Epstein asked them how much their opinions of the candidate had shifted. He called the effect he was measuring the “Search Engine Manipulation Effect,” or SEME.
Based on the results of this research, Epstein testified before Congress that “biased search results can easily produce shifts in the opinions and voting preferences of undecided voters by up to 80 percent in some demographic groups.”
Epstein has used his SEME results to make other bold and specific claims, alleging that as of 2015, “upwards of 25 percent of the national elections in the world were being decided by Google’s search engine.” To arrive at this number, he compared the actual winning margins in elections around the globe to the proportion of undecided voters’ opinions he was able to shift in his SEME research and “just kind of put the two together.”
“The logic he described is just wrong,” says Aaron Brown, a statistician and former chief risk manager at the hedge fund AQR Capital Management who examined Epstein’s research at our request. Brown says that when calculating that number, Epstein makes the “absurd assumption” that biased search engines are the only thing influencing votes. Brown notes that people’s political opinions are shaped by everything from their family to their religion to an ad they might have seen. Epstein’s studies, says Brown, feature artificial scenarios that remove all the other sources of information and influence participants would normally be exposed to in a real election.
Epstein says he has reproduced SEME in realistic situations, pointing to an online study he ran in 2014 on Indian voters during a national election, which produced shifts in preference of 60 percent or more among some demographics. In that experiment, participants used Epstein’s biased search engine for an average of five minutes and then reported whether their opinions of the candidates had shifted after seeing the results.
“Whether that translates to actual changes in the vote people ultimately cast—I think there’s zero evidence for that,” Brown responds. He says that participants in this type of behavioral research often answer in the way that they think the experimenter wants them to; even if their opinions really do change, by the time they show up at an actual polling site, it’s likely that all sorts of other factors have changed their opinions yet again.
“The people whose votes you’re going to shift with something as trivial as five minutes on a biased search engine are going to be the people whose votes are easiest to shift,” says Brown, adding that they’re also probably the least likely to vote at all. To show that SEME was actually influencing elections, Brown says that Epstein needed to have measured “how many actual votes shifted, not how many people reported a shift in opinion.”
As evidence that tech companies can affect real voting behavior, Epstein points to a study run on 61 million Facebook users on election day 2010 that tested the impact of get-out-the-vote messages. He emphasizes that the study found that an additional 340,000 people voted as a result of all the manipulations the experimenters tested.
But Brown says this fits Epstein’s pattern of focusing on seemingly large numbers without giving their proper context. The Facebook study found that only one out of five people who said they voted after seeing the get-out-the-vote message actually did. The most effective appeal increased real voting by just 0.4 percent.
Brown concludes that the Facebook study “makes Epstein’s higher estimates of his effects seem implausible,” and he suggests that far fewer people would actually change their votes as compared to the effects Epstein has observed. He says that because the experimenters measured actual voting, the results are “not a theoretical thing based on a sort of chain of logic.”
Asked to respond to Brown’s criticisms, Epstein says his estimate that tech companies could shift 15 million votes this year is “conservative, based on what I know we can do in our experiments.” And though his studies measured shifts in opinion and not voting behavior directly, Epstein says that “research on polling suggests that people’s stated intentions are excellent predictors of their votes.”
After his SEME experiments, Epstein still had to answer one more critical question before he could arrive at the number of real votes Google was changing: Is the company actually biasing the information it shows to users?
Epstein says Google and other tech companies are “absolutely determined to control the outcome of the upcoming election,” pointing to leaked documents published by the conservative activist group Project Veritas and interviews with ex-employees alleging that Google sometimes intervenes to rerank search results or remove content it deems objectionable.
Tech platforms have often been inconsistent in explaining how they apply their community guidelines to remove or limit certain content. But experts in online search argue the technology is necessarily complex and always evolving. Several studies have concluded that there’s no clear evidence Google biases results for political purposes. Company executives also deny ever doing so.
Epstein needed more evidence that they were doing just that. So, in 2016, he ran a study in which he secretly monitored the search results of 95 “field agents” in 24 states after they typed neutral terms, such as “Hillary’s health plan,” into Google, Bing, and Yahoo. He found a pro–Hillary Clinton bias in all 10 positions on the first page of Google’s results, but not on Bing or Yahoo. After measuring the bias, Epstein combined these results with the shifts in opinion he was able to generate in his earlier SEME research to calculate how many votes Google had supposedly shifted in real national elections.
But Epstein wasn’t able to determine the source of the bias he measured, and his monitoring work has been criticized by other social scientists, in part for using a small number of participants who are unlikely to be representative of the U.S. electorate. And though his claims focus on undecided voters, his data on the bias they’re exposed to are especially limited. Of his 95 field agents in 2016, only 21 identified themselves as undecided, and he didn’t ask the political preferences of others.
Epstein emphasizes that the monitoring systems he’s set up thus far have been “proof-of-concept projects.” Going forward, he says “we want to have numbers that aren’t just statistically valid—we want to have numbers that are psychologically valid,” which would mean collecting “literally millions of pieces of data coming in every day from a very diverse group of people whose demographics we know.”
But he doesn’t regret extrapolating directly from the data he’s collected thus far and widely publicizing the resulting numbers. In his testimony before the Senate Judiciary Committee, Epstein called his estimate that Google shifted at least 2.6 million votes to Hillary Clinton in 2016 “a rock bottom minimum—the range is between 2.6 and 10.4 million.”
“I’m very careful to word things in a way that I think has integrity,” Epstein tells Reason, asserting that he’s never claimed that the bias he observed in his small sample group is indicative of the experience of all Americans. “I’ve never said I know that level of bias was present nationwide,” he emphasizes. “I’ve never said that.”
Brown says the issue is not Epstein’s experimental design, but the fact that he continues to repeat numbers that would only be accurate if the effects he’s measured were the sole factors influencing voters. In the context of all the things that influence votes, Brown says that Epstein hasn’t convinced him SEME is a significant one.
In Epstein’s reply to Brown’s critique, he writes that Brown “has no idea how powerful new forms of influence can be—especially influence that people can’t see.” Epstein argues that tech companies’ influence is more insidious than other forms of attempted online voter manipulation, such as political ads, biased news stories, or Russian troll farms, because users assume computer-generated content is neutral, and because they can be manipulated repeatedly over long periods of time without knowing it.
Epstein believes that the influence he’s measuring is just the tip of the iceberg, since tech companies have vast troves of user-specific data to mine and can use many other so-called ephemeral effects—through search and video suggestions, answer boxes, newsfeeds, and more—to target and manipulate voters in the real world on an ongoing basis.
But Brown says the world is full of implicit messaging, and, unlike in Epstein’s research, studies have shown directly that other types of unconscious influence can have a large effect on real voting behavior. He cites as an example the fact that the top person on the ballot gets between one and 10 percent more votes than they would get if the ballot order were reversed.
“There’s plenty of subliminal and implicit messages that people get all the time,” says Brown. “You don’t need an international conspiracy or a secret group of technocrats in Silicon Valley to influence elections.”
Epstein has argued in the past that Google’s business model should be prohibited by law and that the search engine should be regulated like a phone company. He now advocates for what he calls “light-touch regulation,” arguing that the company’s search index, or the database of web content on which it builds search results, should be made into a public commons. He points to the E.U. for guidance on how aggressively the U.S. government should police tech platforms.
The irony is that some tech CEOs actually welcome regulation, because they already have teams of attorneys and lobbyists at their disposal but new legal requirements are costly barriers to entry for upstart competitors. A study done one year after the E.U. passed sweeping regulations in 2018 in the name of protecting privacy found that Google and Facebook had in fact gained market share.
Epstein’s research is again gaining traction this election cycle among pundits and politicians eager for evidence that Big Tech has it out for the GOP, and his alarming estimates of the real votes these companies are shifting have helped politicians make the case for regulating the internet.
Trump has cited “rigged search results” in calling for greater federal control over online speech, and he recently signed an executive order targeting social media. In July, he tasked the Federal Communications Commission with investigating tech platforms for violations of Section 230, the law sometimes known as the internet’s First Amendment. Republicans in Congress have proposed legislation that would largely eliminate the speech protections in Section 230, allowing the government to punish tech companies if officials were to deem a user’s post unacceptable, and the Department of Justice is reportedly planning to file a new antitrust case against Google before Election Day.
And it’s not just Republicans agitating for control over online speech. Senate Democrats joined them in a bipartisan effort to subpoena tech CEOs in advance of the election. Numerous prominent Democrats, including presidential nominee Joe Biden and running mate Kamala Harris, have promised further crackdowns on internet platforms for the content that users post.
Heading into the 2020 vote, Epstein says his only interest is in free and fair elections. But his alarmist predictions have aided politicians’ opportunistic efforts to get Americans to hand them more power over the internet and to control the free exchange of ideas online. Despite the limitations of his research, Epstein continues to speculate on the number of votes tech companies are shifting, confident that his claims haven’t gone nearly far enough.
Produced, written, and edited by Justin Monticello; camera by Zach Weissmueller and John Osterhoudt; graphics by Austin Bragg, Lex Villena, and Paul Detrick; audio production by Ian Keyser.
Music: Cooper Cannell, Futuremono, Kyle Preston, Quincas Moreira, and Lex Villena.
Photos: 190360221 © Martyn Scates | Dreamstime.com; Photo 99491079 © Alex Grichenko | Dreamstime.com; Illustration 151893782 © Siarhei Nosyreu | Dreamstime.com; Photo 4031959 © Momentsintime | Dreamstime.com; Illustration 183296338 © Skypixel | Dreamstime.com; Illustration 147020235 © Cgracer | Dreamstime.com; Photo 131140532 © Tero Vesalainen | Dreamstime.com; ID 4949445 © Eti Swinford | Dreamstime.com; ID 143609625 © Palinchak | Dreamstime; Pavlo Gonchar/ZUMA Press/Newscom; ID 137038247 © Ian Mcglasham| Dreamstime.com