Facebook Data Proving It Encourages Divisiveness Was Buried By Its Leaders

To Facebook's leaders, divisiveness is a feature, not a bug. It enhances engagement, which drives advertising, leading to financial and political power.

The question now is whether the combination of misinformation, tolerance of disinformation and destructive impact will finally be enough for concerned citizens and governments to put an end to the abuses. JL

Jeff Horwitz and Deepa Seetharaman report in the Wall Street Journal:

Facebook is under fire for making the world more divided. Many of its own experts appeared to agree. “Our algorithms exploit the human brain’s attraction to divisiveness,” read a 2018 presentation. “If left unchecked,” Facebook would feed users “more and more divisive content in an effort to gain user attention & increase time on the platform.” Accounts with hyperactive engagement were far more partisan than normal. (But) Zuckerberg and other senior executives argued that efforts to make conversations on the platform more civil would disproportionately affect conservative users and publishers.

A Facebook Inc. FB -0.16% team had a blunt message for senior executives. The company’s algorithms weren’t bringing people together. They were driving people apart.

“Our algorithms exploit the human brain’s attraction to divisiveness,” read a slide from a 2018 presentation. “If left unchecked,” it warned, Facebook would feed users “more and more divisive content in an effort to gain user attention & increase time on the platform.”

That presentation went to the heart of a question dogging Facebook almost since its founding: Does its platform aggravate polarization and tribal behavior?

The answer it found, in some cases, was yes.

Facebook had kicked off an internal effort to understand how its platform shaped user behavior and how the company might address potential harms. Chief Executive Mark Zuckerberg had in public and private expressed concern about “sensationalism and polarization.”

But in the end, Facebook’s interest was fleeting. Mr. Zuckerberg and other senior executives largely shelved the basic research, according to previously unreported internal documents and people familiar with the effort, and weakened or blocked efforts to apply its conclusions to Facebook products.

Facebook policy chief Joel Kaplan, who played a central role in vetting proposed changes, argued at the time that efforts to make conversations on the platform more civil were “paternalistic,” said people familiar with his comments.

Another concern, they and others said, was that some proposed changes would have disproportionately affected conservative users and publishers, at a time when the company faced accusations from the right of political bias.

Facebook revealed few details about the effort and has divulged little about what became of it. In 2020, the questions the effort sought to address are even more acute, as a charged presidential election looms and Facebook has been a conduit for conspiracy theories and partisan sparring about the coronavirus pandemic.

In essence, Facebook is under fire for making the world more divided. Many of its own experts appeared to agree—and to believe Facebook could mitigate many of the problems. The company chose not to.

Mr. Kaplan in a recent interview said he and other executives had approved certain changes meant to improve civic discussion. In other cases where proposals were blocked, he said, he was trying to “instill some discipline, rigor and responsibility into the process” as he vetted the effectiveness and potential unintended consequences of changes to how the platform operated.

Internally, the vetting process earned a nickname: “Eat Your Veggies.”

Americans were drifting apart on fundamental societal issues well before the creation of social media, decades of Pew Research Center surveys have shown. But 60% of Americans think the country’s biggest tech companies are helping further divide the country, while only 11% believe they are uniting it, according to a Gallup-Knight survey in March.

At Facebook, “There was this soul-searching period after 2016 that seemed to me this period of really sincere, ‘Oh man, what if we really did mess up the world?’ ” said Eli Pariser, co-director of Civic Signals, a project that aims to build healthier digital spaces, and who has spoken to Facebook officials about polarization.

Mr. Pariser said that started to change after March 2018, when Facebook got in hot water after disclosing that Cambridge Analytica, the political-analytics startup, improperly obtained Facebook data about tens of millions of people. The shift has gained momentum since, he said: “The internal pendulum swung really hard to ‘the media hates us no matter what we do, so let’s just batten down the hatches.’ ”

In a sign of how far the company has moved, Mr. Zuckerberg in January said he would stand up “against those who say that new types of communities forming on social media are dividing us.” People who have heard him speak privately said he argues social media bears little responsibility for polarization.

He argues the platform is in fact a guardian of free speech, even when the content is objectionable—a position that drove Facebook’s decision not to fact-check political advertising ahead of the 2020 election.

‘Integrity Teams’

Facebook launched its research on divisive content and behavior at a moment when it was grappling with whether its mission to “connect the world” was good for society.

Fixing the polarization problem would be difficult, requiring Facebook to rethink some of its core products. Most notably, the project forced Facebook to consider how it prioritized “user engagement”—a metric involving time spent, likes, shares and comments that for years had been the lodestar of its system.

Championed by Chris Cox, Facebook’s chief product officer at the time and a top deputy to Mr. Zuckerberg, the work was carried out over much of 2017 and 2018 by engineers and researchers assigned to a cross-jurisdictional task force dubbed “Common Ground” and employees in newly created “Integrity Teams” embedded around the company.

Even before the teams’ 2017 creation, Facebook researchers had found signs of trouble. A 2016 presentation that names as author a Facebook researcher and sociologist, Monica Lee, found extremist content thriving in more than one-third of large German political groups on the platform. Swamped with racist, conspiracy-minded and pro-Russian content, the groups were disproportionately influenced by a subset of hyperactive users, the presentation notes. Most of them were private or secret.

The high number of extremist groups was concerning, the presentation says. Worse was Facebook’s realization that its algorithms were responsible for their growth. The 2016 presentation states that “64% of all extremist group joins are due to our recommendation tools” and that most of the activity came from the platform’s “Groups You Should Join” and “Discover” algorithms: “Our recommendation systems grow the problem.”

Ms. Lee, who remains at Facebook, didn’t respond to inquiries. Facebook declined to respond to questions about how it addressed the problem in the presentation, which other employees said weren’t unique to Germany or the Groups product. In a presentation at an international security conference in February, Mr. Zuckerberg said the company tries not to recommend groups that break its rules or are polarizing.

“We’ve learned a lot since 2016 and are not the same company today,” a Facebook spokeswoman said. “We’ve built a robust integrity team, strengthened our policies and practices to limit harmful content, and used research to understand our platform’s impact on society so we continue to improve.” Facebook in February announced $2 million in funding for independent research proposals on polarization.

The Common Ground team sought to tackle the polarization problem directly, said people familiar with the team. Data scientists involved with the effort found some interest groups—often hobby-based groups with no explicit ideological alignment—brought people from different backgrounds together constructively. Other groups appeared to incubate impulses to fight, spread falsehoods or demonize a population of outsiders.

In keeping with Facebook’s commitment to neutrality, the teams decided Facebook shouldn’t police people’s opinions, stop conflict on the platform, or prevent people from forming communities. The vilification of one’s opponents was the problem, according to one internal document from the team.

“We’re explicitly not going to build products that attempt to change people’s beliefs,” one 2018 document states. “We’re focused on products that increase empathy, understanding, and humanization of the ‘other side.’ ”

Hot-button issues

One proposal sought to salvage conversations in groups derailed by hot-button issues, according to the people familiar with the team and internal documents. If two members of a Facebook group devoted to parenting fought about vaccinations, the moderators could establish a temporary subgroup to host the argument or limit the frequency of posting on the topic to avoid a public flame war.

Another idea, documents show, was to tweak recommendation algorithms to suggest a wider range of Facebook groups than people would ordinarily encounter.

Building these features and combating polarization might come at a cost of lower engagement, the Common Ground team warned in a mid-2018 document, describing some of its own proposals as “antigrowth” and requiring Facebook to “take a moral stance.”

Taking action would require Facebook to form partnerships with academics and nonprofits to give credibility to changes affecting public conversation, the document says. This was becoming difficult as the company slogged through controversies after the 2016 presidential election.

“People don’t trust us,” said a presentation created in the summer of 2018.

The engineers and data scientists on Facebook’s Integrity Teams—chief among them, scientists who worked on newsfeed, the stream of posts and photos that greet users when they visit Facebook—arrived at the polarization problem indirectly, according to people familiar with the teams. Asked to combat fake news, spam, clickbait and inauthentic users, the employees looked for ways to diminish the reach of such ills. One early discovery: Bad behavior came disproportionately from a small pool of hyperpartisan users.

A second finding in the U.S. saw a larger infrastructure of accounts and publishers on the far right than on the far left. Outside observers were documenting the same phenomenon. The gap meant even seemingly apolitical actions such as reducing the spread of clickbait headlines—along the lines of “You Won’t Believe What Happened Next”—affected conservative speech more than liberal content in aggregate.

That was a tough sell to Mr. Kaplan, said people who heard him discuss Common Ground and Integrity proposals. A former deputy chief of staff to George W. Bush, Mr. Kaplan became more involved in content-ranking decisions after 2016 allegations Facebook had suppressed trending news stories from conservative outlets. An internal review didn’t substantiate the claims of bias, Facebook’s then-general counsel Colin Stretch told Congress, but the damage to Facebook’s reputation among conservatives had been done.

Every significant new integrity-ranking initiative had to seek the approval of not just engineering managers but also representatives of the public policy, legal, marketing and public-relations departments.

Lindsey Shepard, a former Facebook product-marketing director who helped set up the Eat Your Veggies process, said it arose from what she believed were reasonable concerns that overzealous engineers might let their politics influence the platform.

“Engineers that were used to having autonomy maybe over-rotated a bit” after the 2016 election to address Facebook’s perceived flaws, she said. The meetings helped keep that in check. “At the end of the day, if we didn’t reach consensus, we’d frame up the different points of view, and then they’d be raised up to Mark.”

Scuttled projects

Disapproval from Mr. Kaplan’s team or Facebook’s communications department could scuttle a project, said people familiar with the effort. Negative policy-team reviews killed efforts to build a classification system for hyperpolarized content. Likewise, the Eat Your Veggies process shut down efforts to suppress clickbait about politics more than on other topics.

Initiatives that survived were often weakened. Mr. Cox wooed Carlos Gomez Uribe, former head of Netflix Inc.’s recommendation system, to lead the newsfeed Integrity Team in January 2017. Within a few months, Mr. Uribe began pushing to reduce the outsize impact hyperactive users had.

Under Facebook’s engagement-based metrics, a user who likes, shares or comments on 1,500 pieces of content has more influence on the platform and its algorithms than one who interacts with just 15 posts, allowing “super-sharers” to drown out less-active users. Accounts with hyperactive engagement were far more partisan on average than normal Facebook users, and they were more likely to behave suspiciously, sometimes appearing on the platform as much as 20 hours a day and engaging in spam-like behavior. The behavior suggested some were either people working in shifts or bots.

One proposal Mr. Uribe’s team championed, called “Sparing Sharing,” would have reduced the spread of content disproportionately favored by hyperactive users, according to people familiar with it. Its effects would be heaviest on content favored by users on the far right and left. Middle-of-the-road users would gain influence.

Mr. Uribe called it “the happy face,” said some of the people. Facebook’s data scientists believed it could bolster the platform’s defenses against spam and coordinated manipulation efforts of the sort Russia undertook during the 2016 election.

Mr. Kaplan and other senior Facebook executives pushed back on the grounds it might harm a hypothetical Girl Scout troop, said people familiar with his comments. Suppose, Mr. Kaplan asked them, that the girls became Facebook super-sharers to promote cookies? Mitigating the reach of the platform’s most dedicated users would unfairly thwart them, he said.

Mr. Kaplan in the recent interview said he didn’t remember raising the Girl Scout example but was concerned about the effect on publishers who happened to have enthusiastic followings.

The debate got kicked up to Mr. Zuckerberg, who heard out both sides in a short meeting, said people briefed on it. His response: Do it, but cut the weighting by 80%. Mr. Zuckerberg also signaled he was losing interest in the effort to recalibrate the platform in the name of social good, they said, asking that they not bring him something like that again.

Mr. Uribe left Facebook and the tech industry within the year. He declined to discuss his work at Facebook in detail but confirmed his advocacy for the Sparing Sharing proposal. He said he left Facebook because of his frustration with company executives and their narrow focus on how integrity changes would affect American politics. While proposals like his did disproportionately affect conservatives in the U.S., he said, in other countries the opposite was true.

Other projects met Sparing Sharing’s fate: weakened, not killed. Partial victories included efforts to promote news stories garnering engagement from a broad user base, not just partisans, and penalties for publishers that repeatedly shared false news or directed users to ad-choked pages.

The tug of war was resolved in part by the growing furor over the Cambridge Analytica scandal. In a September 2018 reorganization of Facebook’s newsfeed team, managers told employees the company’s priorities were shifting “away from societal good to individual value,” said people present for the discussion. If users wanted to routinely view or post hostile content about groups they didn’t like, Facebook wouldn’t suppress it if the content didn’t specifically violate the company’s rules.

Mr. Cox left the company several months later after disagreements regarding Facebook’s pivot toward private encrypted messaging. He hadn’t won most fights he had engaged in on integrity ranking and Common Ground product changes, people involved in the effort said, and his departure left the remaining staffers working on such projects without a high-level advocate.

The Common Ground team disbanded. The Integrity Teams still exist, though many senior staffers left the company or headed to Facebook’s Instagram platform.

Mr. Zuckerberg announced in 2019 that Facebook would take down content violating specific standards but where possible take a hands-off approach to policing material not clearly violating its standards.

“You can’t impose tolerance top-down,” he said in an October speech at Georgetown University. “It has to come from people opening up, sharing experiences, and developing a shared story for society that we all feel we’re a part of. That’s how we make progress together.”