No, Google Did Not Steal The Election

Image by C.E. Kent

So Donald Trump claimed yesterday that Google stole the election:

Wow, Report Just Out! Google manipulated from 2.6 million to 16 million votes for Hillary Clinton in 2016 Election! This was put out by a Clinton supporter, not a Trump Supporter! Google should be sued. My victory was even bigger than thought! @JudicialWatch

— Donald J. Trump (@realDonaldTrump) August 19, 2019

Fact-checkers have already been all over this. Politifact, CNN and WaPo all have fact checkers, disputing this claim. Among other things, Trump is mischaracterizing the research, which claimed that 2.6-10.4 million voters may have been influenced by Google search results to vote for Clinton instead of Trump. The research doesn’t claim that Google manipulated voters because … well, Google can’t do that.

Those are all good fact checks and I recommend them. But I don’t think any of them really get to the heart of the problem here. The problem is one of extrapolation. The claim comes from psychologist Robert Epstein, who has testified before Congress on the subject. Epstein’s chain of logic goeth thusly:

First, he and his researchers have previously claimed that search engine results can have a significant effect on elections, swinging undecided voters by as much as 20% or more. Four of their experiments involved asking volunteers in the United States to judge candidates in Australia’s election based on search engine results, with the search engines canted to favor one candidate or the other. The other involved having voters in India’s 2014 election give their impressions of candidates before and after using a search engine canted to give biased results. “Biased” in both studies meant preferentially linking to articles critical of one or the other candidate. They found that after using the biased search engines, the voters were more likely to favor the candidate with more positive search engine results.

While the study is kind of interesting on its own, using it as a basis to extrapolate anything in the real world is, to put it mildly, insanity. The number of San Diego voters who know anything about Australian politics is effectively nil. Google search results would be literally the first thing they had ever learned about the candidates. The India study is a bit more relevant in that respect, but it also suffers from the same fundamental problem: it’s not clear how those impressions translate into actual votes. Someone’s vote in an election can be affected by many things: news stories, video clips, debate performances, conversations with friends. Claiming that a fleeting impressions based on a search engine propagates to a vote in a predictable manner is suspect, to say the least.

Moreover, this says nothing about the millions and millions of voters who never use a search engine to learn about the candidates. I’m a political junkie; I was blogging every day during the 2016 election. I’m not sure I Googled “Donald Trump” or “Hillary Clinton” even once. Neither of my parents is a political junkie but they vote. And I doubt they Googled either candidate. Most of what they learned, they learned from television news (Fox News in one case; BBC America in the other).

And even putting those objections aside, it’s still not clear at all what this says about the 2016 election. India’s election ran on a time scale of about six months. Opinion polling didn’t even start until three months before the election began. Australia’s election took place on an even shorter time scale. In both cases, most people were only roughly familiar with the candidates. You simply can not compare that to the grueling death march that was the 2016 US election, which ran for almost two years and involved two of the most famous people on the planet. Very few people had to Google Donald Trump to know who he was or what he stood for.

Let me be clear: it’s not that the research isn’t interesting; it’s that it has no practical application to the real world and certainly not to the 2016 election.

The second link in this chain is a study published only on a website. Based on a sample of 95 volunteers, they found that Google’s search results in the run-up to the election favored more negative coverage of Trump than Clinton based on crowd-sourcing the analysis to… a bunch of people. But this claim is also incredibly dubious. Google’s search results are personalized and can vary dramatically from voter to voter. Epstein’s sample was hardly representative, having a larger number of undecided voters (20 people, about 21%) than the overall electorate.

Moreover, what exactly does bias mean? Users ranked sites based on negative or positive coverage. But that’s not necessarily bias. Epstein’s survey covered the 25 days before the election. If a candidate were having a negative news cycle — like, say, an audio tape of him boasting about sexually assaulting women had broken just before the study began — you would expect the coverage to be “biased”. Epstein finds it suspicious that the bias dropped after the election. But that makes sense because the news coverage shifted from horserace stories about Trump’s numerous scandals to stories about the transition and the negative reaction of Clinton supporters to the result.

Google search results can also be “biased” because the most pro-Trump sites were opinion sites like Breitbart, not news sites. Notably, Epstein finds that the “bias” increases as one goes down the list of search results. But what is that but a shifting from news sites to opinion sites? The first hit on the Google searches doesn’t appear to have been “biased” at all.

And, once again, there is no way of knowing how this translates into actual votes. There’s no way of knowing whether those undecided voters actually changed their votes based on search results. Or how many voters were even using search results.

Which brings us finally to the third link in this rusted chain of unreason. If you take the results of the first study and multiply it by the second, you get 5% of the votes being swung by search results. But multiplying two dubious studies by each other does not produce good social science; it produces junk science because any of the objections raised in this entire chain of reasoning can sink the entire result. The claim that so many votes were swung is based on a giant pile of assumptions, a cat’s cradle of humbug. And pulling any thread in the cradle causes the entire thing to unravel.

What if you can’t extrapolate Americans’ opinions of Aussie election to American’s opinions of American elections? What if you can’t extrapolate Indian elections to American ones? What if people’s impressions after a Google search are disconnected from their final votes? What if your sample of 95 people — NINETY-FIVE PEOPLE OUT OF 130 MILLION VOTERS — was biased in some way? What if the people who analyzed the bias in search engine results were themselves biased? What if reality has a liberal bias because Trump had a huge scandal break right as the study began? Each and every one of these questions introduces uncertainties far greater than the plus-or-minus five million Epstein is claiming. He’s telling us to ignore the raging sea and concentrate on a glass of water.

This is one the classic methods of junk science — multiplying humbug by humbug, extrapolating upward and onward until you get dramatic conclusions that defy any common sense (I once addressed, in depth, a similar chain of unreasoning that claimed that rape pregnancies were rare). You take assumption after assumption, bias after bias until all you’ve got is noise that you insist is a signal. Note the absurd precision: 2.6 million is the “rock bottom” number of votes that Google influenced. No responsible social scientist would ever make such a precise claim. Certainly not based on … again … 95 voters.

Of course, this goes beyond the social science. It’s no accident that the Republicans and Trump have been all over this at a time when they are both targeting tech companies for heavy-handed, even tyrannical regulation and insisting that Democrats are trying to steal elections. I think Trump sees this bogus study as two-fer: he can claim political bias in tech companies and delegitimize the 2020 election in advance.

It is very dangerous territory for a president to be treading on. And everyone — Republican or Democrat — should be pushing back on it as hard as possible. Our political system depends on trust. We have to trust that electoral results are valid. If Trump succeeds in destroying that trust, the collapse of the system will not be far behind.

I don’t think Trump is deliberately undermining electoral legitimacy as some sort of grand conspiracy. I think this is mainly to assuage his giant ego. Trump has a history of jumping on garbage claims to pretend he actually won the popular vote in 2016, the loss of which clearly sticks in his craw. But just like a child doesn’t intend to break the glasses when they climb onto a cabinet, his intention doesn’t really matter. What matters is the damage he will do. And making grandiose claims based on dubious science doesn’t help; it merely hands the child a blowtorch.