Starting in August 2019, you may have seen an ad in your Facebook news feed asking you to take a news quiz. If you didn’t know who controlled the Senate, for instance—about 30% of people didn’t—you would be classified as most persuadable, and you would become part of one of the largest and most sophisticated experiments of its kind.

On the internet, we’re subject to hidden A/B tests all the time, but this one was also part of a political weapon: a multimillion-dollar tool kit built by a team of Facebook vets, data nerds, and computational social scientists determined to defeat Donald Trump. The goal is to use microtargeted ads, follow-up surveys, and an unparalleled data set to win over key electorates in a few critical states: the low-education voters who unexpectedly came out in droves or stayed home last time, the voters who could decide another monumental election.

By this spring, the project, code named Barometer, appeared to be paying off. During a two-month period, the data scientists found that showing certain Facebook ads to certain possible Trump voters lowered their approval of the president by 3.6%. For the frantic final laps, they’ve set their sights on motivating another key group of swing-state voters—young Democratic-leaning voters, mostly women and people of color—who could push Joe Biden to victory.

“We’ve been able to really understand how to communicate with folks who have lower levels of political knowledge, who tend to be ignored by the political process,” says James Barnes, a data and ads expert at the all-digital progressive nonprofit Acronym, who helped build Barometer. This is familiar territory: Barnes spent years on Facebook’s ads team, and in 2016 was the “embed” who helped the Trump campaign take Facebook by storm. Last year, he left Facebook and resolved to use his battle-tested tactics to take down his former client.

“We have found ways to find the right news to put in front of them, and we found ways to understand what works and doesn’t,” Barnes says. “And if you combine all those things together, you get a really effective approach, and that’s what we’re doing.”

As the pandemic has glued voters to their phones and screens and added unprecedented hurdles to campaigning and voting, the 2020 digital arms race has become bigger and uglier than many had expected. The presidential and congressional campaigns have spent more than $1 billion on digital ads this year, according to the Center for Responsive Politics, during what’s become the costliest election in history. In the final weeks, the online ad shopping spree has grown so intense that in many battleground states, YouTube ads are sold out.

The president had a running start. On top of what his former campaign manager called a “Death Star” of data and campaign infrastructure, the Trump campaign has devoted far more of its resources to digital advertising than the Biden effort, and spent a whopping $142 million on Facebook ads so far, outpacing Biden’s outlay of $89 million, according to an analysis of the Facebook ad archive.

A multitude of outside anti-Trump groups such as Acronym have spent millions more to fill in the gaps. Earlier this year, Priorities USA and Color of Change launched a $24 million digital advertising campaign aimed at exciting Black voters in swing states. American Bridge and Unite the Country, two of the other largest progressive PACs, have tapped Mike Bloomberg’s political ad tech startup, Hawkfish to wage their own data-rich digital onslaughts through Election Day. Acronym was first out of the gate, and is thought to be the Democrats’ most advanced digital advertising project. By the election it promises to have spent $75 million on Facebook, Google, Instagram, Snapchat, Hulu, Roku, Viacom, Pandora, and anywhere else valuable voters might be found.

For a year that money went toward targeting low-information voters in Michigan, Pennsylvania, Wisconsin, Arizona, and North Carolina, but by the end of the summer, the Barometer team saw its persuasion powers diminishing; they guessed that they couldn’t budge the president’s approval rating any lower. So Acronym redirected that cash to motivate another critical audience of low-information voters: new or unlikely Democratic-leaning people thought to be unexcited about Biden and his running mate, Senator Kamala Harris. Barometer’s scientists have identified 1.8 million such voters in six states—mostly women of color younger than 35 across Acronym’s original five target states, plus Georgia.

With more than $1 million per week in Facebook ads during the homestretch, “we’re trying to boost their enthusiasm,” says Kyle Tharp, Acronym’s VP of communications.

Despite upbeat polls and record early turnout numbers, Acronym’s battle was never going to be easy. These voters are thought to be some of the least-excited, and while Acronym has identified them as the easiest to persuade, they are also highly susceptible to the sort of BS that can keep voters home. Research has shown that low-information voters are not only less likely to vote but more likely to believe falsehoods; sometimes they’re called “misinformation voters.” And deterring voters with falsehoods and fear may be easier than motivating them with facts and hope. A false claim about voting, for instance, is much easier to spread on Twitter—or by anonymous text message—than it is to correct.

And there’s no end to the BS. While Team Trump floods swing states with anti-Biden ads meant to dampen enthusiasm among Democrats (attacks on his and Harris’s record on criminal justice reform, for instance), the candidate and many others are waging hybrid war, pushing extreme rumors of child-trafficking cabals and falsehoods about voter fraud and mail-in ballots—a handy way to erode the confidence and will of already hard-to-motivate voters.

“We have three major voter suppression operations underway,” a senior Trump campaign official told Bloomberg in the days before the 2016 election, referring to a barrage of disappearing “dark posts” that painted Hillary Clinton as a racist and targeted white liberals, young women, and African Americans. The Russians took a more direct approach: After identifying groups of people of color using Facebook’s tools, one Russian-backed page microtargeted them with an ad on Election Day: “No one represents Black people. Don’t go to vote.”

pic.twitter.com/eOzQKyVgrF

— PACRONYM (@PACRONYM) October 30, 2020

These are some of the same types of voters that Barnes and his team are trying to motivate. “I think we’re trying to not just fill the gap that other [progressive campaigns] aren’t filling, from a targeting perspective,” he says, “but we’re also trying to find the people who we think are pretty vulnerable to hearing this information and get the facts in front of them.”

Remember the Alamo

Barnes comes with a special skill set and a quiet, fierce sense of urgency. The last time he targeted impressionable voters with relentlessly tested Facebook ads and reams of data was inside Project Alamo, the Trump campaign’s big-data digital operation. Barnes is eager to put 2016 behind him, but readily acknowledges the work he did then was “groundbreaking.” Trump had his Twitter, but no campaign had before used Facebook to persuade voters, find more of them, and raise money like Trump’s had. At times a dollar in ad spend translated to donations of $2 or $3. Facebook’s then-head of ad technology wrote in an internal memo that it was “the single best digital ad campaign I’ve ever seen from any advertiser. . . . They just used the tools we had to show the right creative to each person.” In a since-deleted tweet, Trump’s digital director, called Barnes “our MVP.”

pic.twitter.com/V8At6Xt9D5

— PACRONYM (@PACRONYM) October 30, 2020

In the wake of Trump’s election and its aftermath, Barnes helped Facebook develop some of its election integrity initiatives (one of Facebook’s moves was to stop embedding employees like him inside campaigns) and even sat down for lengthy interviews with the Securities and Exchange Commission and with then-Special Counsel Robert Mueller. Last year, after some soul-searching, some of it in Peru, Barnes registered as a Democrat, left Facebook, and began working on a way to fight Trump. He reached out to Tatenda Musapatike, another former political ads specialist at Facebook, who had joined Acronym as its senior director of campaigns. “I knew that I had an extra set of experiences and talents that could be helpful to help be a part of this,” he says.

Acronym and a political action committee, Pacronym, were founded in 2017 by Democratic strategist Tara McGowan, in an effort to counter Trump’s online spending advantage and what The New Yorker called his Facebook juggernaut. As a nonprofit, Acronym does not disclose its donors, though Pacronym does, and the groups’ $75 million battle plan drew backing from a who’s who of California megadonors: Dollar Shave Club founder Michael Dubin, Laurene Powell Jobs, Steven Spielberg, and venture capitalists Reid Hoffman and Michael Moritz. Chris Cox, Facebook’s former product officer, is also a donor and informal adviser, according to the Wall Street Journal.

Barnes has used their cash to build an eight-person ad tech dream team: vets of Silicon Valley along with experts in computational social science and machine learning. One of Barnes’s recruits from Facebook, data scientist Solomon Messing, is an expert in social media and politics: At Facebook, one experiment he ran ahead of the 2012 election appeared to show that promoting more news stories in users’ news feeds increased their political awareness and their turnout at the polls. Similar Facebook experiments have shown that Election Day reminders from friends can also push people to the polls; this year, Facebook says its News Feed reminders have led 4.4 million Americans to register to vote.

In some ways, Barnes and his team have rebuilt Brand Lift, an ad-optimization feature on Facebook that Barnes helped the Trump campaign leverage at unprecedented scale. (As part of its post-2016 election integrity reforms, Facebook removed the feature for political advertisers.) By sending follow-up surveys to users who have seen its ads, typically a week after the fact, the Barometer team has collected data on which content, narratives, and messengers are most effective at impacting presidential approval, voter enthusiasm, and vote choice, across a range of issues. To persuade potential Trump voters, for instance, Barometer has found the most traction not with partisan news sites, highly produced ads, or ads at all, but with stories from mainstream news sources. “Not news content with spin, not even editorialized news content, just news content about the president,” Barnes says.

Even working remotely, Barnes believes Barometer is light years ahead of the persuasion efforts he helped develop for the Trump campaign. “We were just shooting in the dark in 2016 compared to what we’re doing now,” he says. “I cannot imagine that the Trump campaign is doing anything similar right now.”

Cleaning up Facebook’s mess

For Barnes, Acronym’s aggressive approach to Facebook, and Barometer’s very existence, isn’t just personal, but relates to his former employer: Facebook hasn’t only failed to effectively police misinformation and disinformation, but helped accelerate it.

“One of the reasons that the work that we’re doing is so important, just reminding low-information voters—folks whom we have a lot of evidence that their opinions are easier to shift, and putting mainstream, fact-checked news in front of folks that help them understand what’s actually going on in the world—is because all of the decisions that Facebook has made, especially in the past few years.”

We knew Trump would try to lie, cheat, and steal this election. That’s what he’s done his whole life. But what none of us could have expected is just how far Facebook would go to help him. pic.twitter.com/u9TCsD0Efk

— PACRONYM (@PACRONYM) October 28, 2020

Barnes is still bothered by Facebook’s 2018 decision to place Breitbart in a section devoted to “deeply reported and well-sourced” journalism, as part of an effort to represent what Mark Zuckerberg called “a diversity of views.” On Wikipedia, Breitbart has been placed on a list of unreliable sources not to be cited, due to instances of inaccurate and incendiary reporting.

“The evolution of the platform has encouraged, over a very long time—I was there obviously for a while—just the vast proliferation of questionably true, if not just outright false information from right-wing publishers, who are often not just prioritized on the feed, but are given space on the news tab,” he says.

While Facebook has come under fire for declining to place fact-checking labels on misleading campaign ads, organic content from dubious right-wing pages and sites like Breitbart, The Daily Wire, and The Federalist have been pumping out their own political falsehoods. At times, Facebook has bent its rules to accommodate popular conservative outlets.

“I think one of the biggest stories of the election is how Facebook has tolerated—and not only just tolerated, but prioritized—the distribution of right-wing misinformation—organically, and not through advertising,” says Tharp. “You know, if Trump is able to pull this out, it will be in no small part due to that sort of stuff spreading on Facebook. And Facebook just shrugs and acts like, you know, ‘to the victors go the spoils,’ ‘they know how to use our platform,’ whatever.”

“So,” adds Barnes, “we’ve got a lot to fight against.”

But while Barnes is using some of the weapons that helped Trump, he’s at pains to emphasize that, unlike the other side, Acronym’s artillery is simply “the facts.” “One of the reasons I’m so proud of the program that we’ve run and so proud of the team and all the tools is there’s nothing misinforming about what we’re doing. We wouldn’t even need to, because we’re just putting the facts in front of people.”

Lessons from the battle

Acronym may say it’s above dirty tricks, but its aggressive approach has still raised eyebrows. Last year it drew fire for launching a network of news sites in battleground states called Courier Newsroom, which it launched in part to counterbalance an ecosystem of right-wing journalism outfits. NewsGuard, a journalism watchdog, said the websites are “insufficiently transparent” and “cherry-pick facts to advance a Democratic narrative.” But while Courier Newsroom is partisan, it does not, as some have charged, publish false stories.

Related: Five ads that could persuade your pro-Trump uncle to vote Democrat

Even as it pushes out its own news sites’ content on Facebook, Acronym has focused heavily on boosted news from more trusted sources, per Barometer’s recommendations. Among older possible Trump voters, Barometer has also found that content from conservative commentators can be particularly persuasive. For instance, when the team messaged its persuasion audience of potential Trump voters with reporting from The Atlantic about the president’s disparaging comments concerning war veterans, they found no meaningful impact on his approval. But a Fox News article confirming Trump’s comments was a more persuasive messenger.

Just as some media are more trusted than others, Acronym’s approach is also informed by the premise that friends are more persuasive messengers than strangers. The idea is core to relational organizing, a tactic pioneered by the Pete Buttigieg campaign that involves getting supporters to message their family, friends, and neighbors to encourage them to vote. Both the Trump and Biden apps ask users to upload their contact lists, which campaigns can then use to identify people who are likely targets, and whom the app users can target with personalized messages.

To reach its persuadable audience of voters through the noise, Acronym and Black PAC, one of its partners, have honed messages that emphasize the gains that Black leaders have made for voters, broad support for the Black Lives Matter movement, and Biden’s and Harris’s history of social and racial justice. Other strategists and super PACs have launched efforts aimed at combating misinformation and voter suppression efforts.

Related: Mistrust of public officials may help Election Day misinformation spread

It’s hard to say how definitive any of Acronym’s findings about voter persuasion are: No one else is running or disclosing experiments like this, Barnes says. And no one else has this kind of data. Whether it all works may be known only after the votes are counted, whenever that is.

After the election, Acronym and Barometer plan to share more of their lessons with the public at large, something they already do frequently with progressive partner organizations. (Law prohibits outside groups from coordinating with the campaigns themselves.) “A big part of our mission is to build power for the progressive movement, and that means making sure research and learnings from this cycle don’t just disappear once the campaign is over,” says Tharp.

What will they do with their digital arsenal in the event of a contested election or other chaos? “We’ve had and continue to have lots of conversations with groups on this,” Tharp says, without going into details.

To prevent last-minute election hijinks, Facebook has banned new political ads for the final week of the U.S. campaigns. And last month, the company appeared to cave to years of pressure when it announced that it would cease selling political ads altogether after the election, for an indefinite period. (Facebook officials have said these ads represent a small portion of its revenue; in the past three months, it reported earning $625 million from political ads, or 7% of its second-quarter U.S. revenue, according to its financial reports and ad archive.)

Freezing political ads, especially in the campaigns’ final week, has ignited frustration on both sides. (On Thursday, Facebook acknowledged that a technical error had caused ads from both political parties to be improperly paused.) An ad ban hobbles the ability of Acronym and other groups to counter political misinformation, says strategist McGowan, who called the final-week ad freeze a “shallow and dangerous public relations move.”

The dark side

On the other hand, part of the reason for Facebook’s ban is to prevent the spread of abusive political microtargeting. McGowan acknowledges the risk. “We have to also note the moral question here, which is: These tools can be used for nefarious purposes,” she said on a recent episode of Acronym’s podcast. “They can be used for bad as well as they can be used for good. . . . And when you have targeting tools and you have measurement tools you can identify how to manipulate people with misinformation in this environment.”

Related: The strange afterlife of Cambridge Analytica and the mysterious fate of its data

In July, a report on political campaigning from the U.K.’s Information Commissioner’s Office called for an “ethical pause” on the use of personal information in political microtargeting so that regulators and industry could reconsider the technology. “These techniques raise fundamental questions about the relationship between privacy and democracy, as concerns about voter surveillance could lead to disengagement with the political process,” U.K. information commissioner Elizabeth Denham wrote.

In Washington, efforts to regulate online ads and personal data have stalled, but a growing bipartisan push to regulate Big Tech could bring new rules. The danger of microtargeting “is not a problem advertising platforms can solve by themselves, as they themselves acknowledge,” a group of social media researchers wrote in a letter last month to the CEOs of Facebook, Google, and Twitter. “The vetting being done by the platforms alone is not working; public transparency of all ads, including ad spend and targeting information, is needed so that advertisers can be held accountable when they mislead or manipulate users.”

Barnes is unimpressed by what he calls bluster about the power of microtargeted “dark arts.” “I think what really was effective in 2016 was just very smart—but not like completely insane voodoo magic—political tactics,” he says, emphasizing that what he and his team have built “is still not nearly as kind of precise and scary—I don’t think it’s scary at all—but it’s not as precise as it’s been described, in the sort of scariest notions of 2016.

“But,” he adds, “it is still quite an advancement from what they did then.”