Inside Google's struggle to filter lies from breaking news
(Bloomberg) -- Facebook has a fake news problem. Google has an evil unicorn problem.
“Evil unicorns” — a term some Google engineers once coined, according to a former executive — are unverified posts on obscure topics, full of lies. They pop up from time to time on the web and find their way into Google's search results. In an ideal world, Google’s search algorithm should force these fake, pernicious creatures so low in search results that they are buried deep in the web where few can find them.
Here's the problem: These unicorns — no, they've got nothing to do with highly valued startups — are designed to surface in a void. And after a breaking news event, like a mass shooting, there’s scant verified information for Google’s engine to promote. As Jonathan Swift once wrote, falsehood flies, and the truth comes limping after it.
"As soon as an event happens, everything is new," said Nate Dame, a search specialist at marketing firm Propecta. "There's no system for the algorithm to filter out truth and reality."
After the Oct. 1 Las Vegas shooting, several accounts seemed to coordinate an effort to smear Geary Danley, a man misidentified as the shooter, with false claims about his political ties. There were no existing web pages or videos broadcasting that Danley was innocent, and in the absence of verified information, Google's algorithms rewarded the lies, placing inaccurate tweets, videos and posts at the top of search results. A month later, when Devin Patrick Kelley shot and killed 26 people in Sutherland Springs, Texas, YouTube videos and tweets mislabeled him as “antifa,” a term for radical, anti-fascist protesters. This was not true, yet Google displayed these posts prominently.
Pandu Nayak, a search senior executive at Alphabet Inc.’s Google, said the newer policies around search "actually worked really well" after the Las Vegas shooting, with the Danley misidentification being a notable exception. "It wasn't this huge problem," Nayak said. "But we should have absolutely anticipated this, but didn't."
This is a familiar headache for the company. For years, Google fought and won a similar battle with spammers, content farms and so-called search engine optimization experts over which web pages should be shown at the top of search results. But these latest web manipulators are causing greater havoc by targeting a slightly different part of Google — its real-time news and video results.
They're exploiting a weakness that cuts to the core of Google's main proposition: Delivering trusted information online. That flaw emerged as Google rewired its search engine and giant video platform to prioritize immediate and timely content to become a destination for news.
"The purveyors of misinformation are really using these methods to complicate our systems,” Nayak said.
To combat the problem, Google is revamping the place where most people first see web results with breaking news, carefully curating the carousels that list “Top Stories” and featured posts Google pulls from Twitter in a way it hasn't before. Nayak said the company is working on methods to limit false content around news events, but declined to offer specifics. Google is also overhauling video search, limiting results around news events on YouTube to verified outlets and placing more algorithmic emphasis on these sources more broadly.
But will these solutions be enough to outsmart the persistent evil unicorns?
"One of the challenges here is that these rumors pick up so fast," Danny Sullivan, Google’s public liaison for search, said. "I can't tell whether there's more of it happening than in the past. It kind of feels like it."
Sullivan spent years as the foremost chronicler, and frequent critic, of Google search, until the company hired him in October. In April, Sullivan wrote a blog post detailing Google's recent stumbles — false election outcomes, results questioning the existence of the Holocaust — calling it the "biggest-ever search quality crisis." Later that month, Google rolled out several reforms meant to address this, including favoring trusted websites, like news outlets, for obscure queries.
Google began its quest to fix search after last year’s U.S. election. The problem was not new. In the past, Google engineers had watched some ugly problems emerge amidst vaccine and climate change controversies. When people convinced of the dangers of childhood vaccines started blogging and posting, there were fewer truthful sources to offset them; doctors hadn’t spent much time blogging about the benefits of vaccines. But the truth was findable online, so Google began to structure search results to assign more weight to authoritative sources. The chaos surrounding news, however, has proven to be a harder challenge.
One reason is that Google has added more real-time information to its search results. In 2014, the company opened up its news results to non-news publications like personal blogs, and a year later Google cut a deal with Twitter to show tweets high in query results as part of a broader effort to turn search into a hub of fresh information and direct answers.
Some critics wonder: Why can’t the company restrict timely results to verified sources?
Google worries that narrowing the pool of websites to trusted sites may cut off the web's niche corners. Nayak gives an example: Fans of a minor hip-hop artist might crave information that only appears on small blogs. "Authoritative sources are not just going to cover all this long-tail of interests that people have," he said.
And vetting news sources is an unwelcome task. Critics have ripped into Google and Facebook for categorizing certain publications, and not others, as news. It’s a political mire Google's search unit is very reluctant to wade into.
Dame and others in the field argue that Google has made this real-time information problem worse by adding more machine learning. These systems, where software is trained to learn on its own, differ from the search algorithms that weigh sites heavily on factors like how many links they’ve received. Because the systems learn from what they have, they’re more adept at fetching a site that's relevant to a given search terms, even if its veracity is unproven.
But machine learning is a crucial tool, Google Chief Executive Officer Sundar Pichai said in an interview in October. "I also think over time, the other actors who are trying to attack your systems will also use machine learning, so I think it's equally important we use machine learning to do more."
In the past year, Google has also teamed with several fact-checking organizations to certify news results. It showed after the Texas shooting. The initial flood of false content about Kelley forced Google's auto-complete function, which suggests searches based on popular queries, to suggest "antifa" as people searched for his name. Yet by Monday, those searches produced top results from fact-check sites, such as Snopes, and other news outlets dispelling the connection. (Of course, a lot fewer people were reading the news by then).
At YouTube, the problems persisted. Search the same term there and, a week later, the first page of results lists videos from CBS News and Fox News along with conspiratorial clips. One was created a few hours after the shooting. A male voice claims he has "one-hundred proof" the shooter is "far-leftist" and "antifa." "If you like this stuff," the voice goes on, "make sure to like or comment or subscribe. I do these things all the time."
He doesn't. The YouTube account, The Patriotic Beast, had posted only two other videos on its channel, which garnered fewer than 500 views. But the Sunday video racked up tens of thousands of views in one day, largely because of a prominent placement in Google search.
Johanna Wright, a YouTube vice president, said the company is working on a sweeping change to query responses. In March, YouTube added a section of "Top News" from verified outlets and placed more of those videos on the YouTube home page. "We saw that wasn't enough," Wright said. Going forward, she said YouTube will lean even more on registered news organizations.
But Wright said YouTube does not want to ditch small-time YouTube posters. It worries about suppressing the work of "citizen journalists." During the Arab Spring, Wright said, many of the people documenting the events on the ground were using recently created YouTube accounts.
And crowding out new and smaller creators isn’t great for business. Google needs to keep YouTube stars from migrating to Facebook, Amazon or elsewhere, and the prospect of a big payoff from search is a large carrot. Matt Jarbo makes YouTube videos for a living, posting about three a day on a range of topics. On the Sunday of the Texas shooting, he recorded a 29-minute video in which he narrates web articles about Kelley. He turned off the ability to run ads on the video, given the topic, but said these type of news-related segments help build his reach for the videos where he does make money.
Besides stuffing the page with key words, Jarbo said there are other tactics for getting attention: Be detailed in descriptions; for video titles and tags, use catchy and colloquial phrases. To go wider on YouTube, he will often track trending topics on Facebook and Google for inspiration. "It seems kind of shady to talk about," he said. "This is what they want. This is the game you have to play."