In the fall of 2018, Jonah Peretti, the chief executive of BuzzFeed, emailed a senior official at Facebook about a list. BuzzFeed had published 21 Things That Almost All White People are Guilty of Saying, and on Facebook the post drew 13,000 shares and 16,000 comments, many of them from people attacking BuzzFeed for writing it and arguing with one another about race. The company's news videos and its pieces on self-care and animals were struggling to be seen at all. His complaint was that the new ranking was not rewarding the meaningful interactions it was named for, and that his staff now felt they had to make worse content or watch their work underperform. The email is known because Facebook's own staff quoted it in internal documents that The Wall Street Journal later reviewed.
MSI was the change Mark Zuckerberg had announced on January 11, 2018. He was changing the goal he gave Facebook's product teams, he wrote, from helping people find relevant content to helping them have more meaningful social interactions, and he expected the time people spent on Facebook, and some measures of engagement, to go down. Inside the company, meaningful had to become a number. Under an internal point system, a like was worth one point and a reaction or a reshare without added text five, while a significant comment, message or reshare earned 30. Further multipliers depended on whether the people interacting were friends, members of the same group or strangers. The News Feed was then arranged to encourage as much of that scoring as possible.
As arithmetic, it worked. Comments, likes and reshares had declined through 2017, and tests in August 2018 found that the change slowed the fall in comments and mostly improved the count of daily active users. A score cannot tell an argument from a conversation, though, and a post that makes people furious collects comments fast. Facebook's data scientists warned in a memo that the approach was having unhealthy effects on politics and news, and one of them later described it as a growing liability. Researchers found misinformation, toxic posts and violent content turning up far out of proportion among reshares, one of the interactions the formula rewarded most.
Political parties noticed. In April 2019, two Facebook researchers reported that one Polish party's social media team estimated it had moved its posts from half positive and half negative to 80 percent negative, and attributed the shift explicitly to the algorithm. Parties in Spain, the researchers wrote, had found that harsh attacks on opponents drew the most engagement, and said that although they tried to resist, they ended up doing whatever worked. The researchers heard similar complaints from parties in Taiwan and India. Their report did not name the parties, and the Polish figure is the party's own estimate as relayed to Facebook, not a measurement the company made.
Facebook had already measured its own systems feeding a worse problem. A 2016 presentation by Monica Lee, a researcher and sociologist at the company, found extremist content thriving in more than a third of large German political groups on the platform. By its count, 64 percent of the joins to extremist groups came through Facebook's own recommendation tools, most of them through the features called Groups You Should Join and Discover, and the company's recommendation systems, it concluded, were making the problem bigger. The figure counts joins, which are a different thing from people, and the Journal's account does not say whether it was measured only on the German groups; other employees said the problem was not confined to Germany. Its findings became public in May 2020, when Jeff Horwitz and Deepa Seetharaman reported them in the Journal. A Facebook spokeswoman answered that the company had learned a lot since 2016 and was not the same company today.
The researchers also proposed remedies. One targeted downstream MSI, the part of the formula that pushed a post higher when the algorithm calculated that people further along a chain of reshares were likely to comment on it or share it again. Early tests found that turning it down for civic and health content reduced the spread of false material, and Facebook made that change in the spring of 2020. That April, Anna Stepanov, who led a team working on the problem, briefed Zuckerberg on extending it more widely. Afterward she wrote to colleagues that Zuckerberg did not think the change could be applied broadly. He was open to testing it, she said, but it would not launch if it meant a material loss in MSI. In August 2021 Facebook announced it was expanding tests that put less emphasis on how likely someone was to comment on or share political content. The Journal published the documents the next month, on September 15, as the third story in a series it called the Facebook Files.
Facebook's reply was that the divisions are older than the feed. A ranking change was not the source of the world's divisions, the spokesman Andy Stone told the Journal, citing research showing partisan divisions growing for decades before platforms like Facebook existed. Outside research later lent that argument support. For a study published in Science in 2023, academics working with Meta moved consenting Facebook and Instagram users onto chronological feeds for three months around the 2020 U.S. election. Those users spent much less time on the platforms. They saw more political and untrustworthy content, and on Facebook less uncivil content, and their levels of polarization did not significantly change.
None of that shows the 2018 change caused polarization. What the documents show is narrower. Facebook named a number after what it wanted, then built its feed to raise the number, and its own researchers kept finding that the two had come apart. The score was an accurate count of comments and reshares. Whether those interactions were meaningful was a separate question, and the people paid to study it wrote down, more than once, that a good deal of what the score rewarded was outrage. A metric counts what it counts, whatever it is called.