Why you should care
Because her A.I.-fueled truth-seeking platform is fighting back against misinformation.
Before immigrating to the Washington, D.C., suburbs as a teenager in 1994, Olya Gurevich watched the Soviet Union collapse. It was thrilling. “Russia went from black and white to an explosion of actual color. Of Western television, clothing, food,” she says. “I came to the United States in this very optimistic state of mind, thinking that all of that was behind us.”
But her faith was shaken two decades later while watching Russian interference in the 2016 election help elect Donald Trump president. Her concerns compelled her to help found Marvelous AI, a San Francisco–based startup examining political speech and social media to sort out fact and fiction in the 2020 presidential race. As one of three cofounders and the chief scientist, Gurevich uses natural language processing to sort through more than 20 million tweets per month — harnessing computational clustering techniques and context clues to examine how online chatter is used to attack or defend a candidate on everything from character to electability.
That information is published online through the Marvelous AI blog and disseminated to reporters. Plus, through a collaboration with Russian state propaganda expert Sarah Oates at the University of Maryland, their findings are used to identify and call out narratives that “are consistent with Russian goals,” says Gurevich, 41.
Take the first Democratic debate. Most pundits, including data luminaries such as Nate Silver, declared Kamala Harris the clear-cut winner for her confrontation with Joe Biden over busing policies and segregation.
But Gurevich’s analysis found something surprising. Yes, Harris had dominated the online conversation. However, 85 percent of the chatter was negative, despite the positive mainstream coverage. In fact, most of the tweets were calling her a liar, a fraud or a sellout. And many — with a conspiratorial tone that suggested nefarious intentions — questioned whether she was actually Black or American (she is both). What’s more, Gurevich’s algorithm also showed that Elizabeth Warren was starting to get steam behind her policy talk, while Pete Buttigieg was fading. In the weeks following the debate, the polls followed Gurevich’s path instead of the pundits’, though the race has shifted again since.
Gurevich’s team also produced a “Running While Female” report on how social media commentary disadvantages women running for president, and Gurevich has been asked to speak at various groups, including the American Political Science Association and Netflix Personalization, Recommendation and Search conferences. Her presence — as a sober-eyed, data-focused surveyor of online narratives — is particularly important as misinformation abounds and the American public loses trust in both polls and the news media.
“There is a technical part of this that is very interesting” both for society and for the scientific field of natural language processing, says Tracy Holloway King, an industry peer and mentor to Gurevich. But there is also a risk: “Especially when you’re in an area where part of the issue is what is true and what is not, if you don’t provide transparency, you yourself could accidentally be feeding into that misinformation.”
We are recognizing that this is the new way of the world. This stuff is not going away.
The pitch for Marvelous AI is simple for the University of California, Berkeley linguistics Ph.D. with past stints at Microsoft and Apple. While there has been a greater emphasis on fact checking since the 2016 election, it’s not enough: Experts need to know what messages are spreading on the internet, who is spreading them and how. Not only politicians should be worried, but all sorts of brands, from companies to universities. “We are recognizing that this is the new way of the world. This stuff is not going away,” Gurevich says.
Marvelous AI wants to eventually become a full-service data company for reputation management. “2020 is our proof of concept,” says Christopher Walker, fellow co-founder and CTO. Gurevich likens it to an anti-virus software “against malware of the cognitive variety.” There are other companies that work to counteract online narratives, although perhaps none that focuses so much on combating political ones. Regardless, Gurevich says they “look forward to partnering with folks who have complementary skills.”
Of course, such a company could easily become a propagandist itself if it helped clients stave off fair criticisms of their products. “We don’t want to be in the business of staging interventions,” Gurevich insists. Walker says the company is primarily using open source coding and plans to make the sources of its analyses as accessible as possible.
The company begins by gathering data — in this case, millions of public Twitter posts (plus the occasional analysis of news articles, blog posts and Google searches) — and uses AI to cluster similar content, allowing a human to look at a few days’ worth of tweets in maybe half an hour. Those tweets are organized into clumps — as negative or positive, as being an attack on policy or personality or something else.
Already, trends have emerged. Warren has often been painted online as a liar or a phony, and Biden has been accused of corruption, both characterizations used to bludgeon Hillary Clinton in 2016. Another pattern? While Twitter users may say similar things (good or bad) about Warren or Bernie Sanders, posts about Biden seem to be almost exactly parroted, word for word. “It’s less authentic activity,” Gurevich says.
That isn’t necessarily a sign of Russian meddling. “It could also mean that the right-leaning Twitter is much more on message than the Democrats,” she says. By comparing social media posts with the news sources users are posting about, Marvelous AI is able to clearly see which “corner of this universe” narratives are emerging from.
Despite all of her work, Gurevich has a surprising end goal. “I’d like for this to be irrelevant,” she says, so that she could spend more time fixing nonpolitical problems, or performing with the San Francisco Bach choir — the soprano is the daughter of a music teacher and sister to an opera pianist — and raising her two daughters. Barring that though, her focus is on building coalitions dedicated to the truth. “If the misinformation in media is here to stay, it’s going to have to be a whole ecosystem of companies, researchers working against them.” A resistance, of sorts.