The Washington PostDemocracy Dies in Darkness

At Facebook, algorithms choose the news. And sometimes they don’t get it quite right.

(Amy Cavenaile/The Washington Post; iStockphoto)

The Intersect ran a little experiment a few weeks ago. During the work day, we’d check in with Facebook each hour, on the hour, and record which topics were trending for us on the platform. The resulting daily pop-up newsletter gave us some interesting insight into the world according to Facebook. We will be exploring some of what we learned in a series of posts in the coming weeks. This is the third in the series; read the first and second here.

A single word change in Facebook Trending can determine whether millions bother to read the news therein.

Would you click “Tulsa, Okla.” or “Tulsa police shooting?”

“Tax Foundation” or “Trump Tax Plan”?

“#HackingHillary” or the more sedate . . . “Hillary Clinton”?

Facebook has repeatedly trended fake news since firing its human editors

All three examples surfaced during the Intersect’s September audit of Trending topics, the much-maligned Facebook feature that attempts to surface the day’s most popular news. During work hours from Aug. 31 to Sept. 21, we logged every news story that trended on Facebook across four separate accounts.

Our findings should be taken as demonstrative, not conclusive, as Facebook personalizes its Trending module to each individual user. Still, we found a range of naming conventions, some of them unusual and all of them destined to influence the reach of their attendant news articles. If only 1 in 20 U.S. Facebook users got news from Trending, the feature would still command an audience of nearly 9.5 million. That’s like the entire city of New York basing its news diet on what Facebook Trending suggests to them.

Generally, we found that Trending topics tend to be represented by proper names — most commonly the name of a person, but also of products, places, organizations and formalized events, as in “Suicide Prevention Week” or “Steelers vs. Redskins.” Those names typically correspond with the primary subject of the news story they’re attached to, but in some cases they do not — particularly, it would seem, when the story relates less to a concrete person or place than to a larger concept.

A viral comic about climate change trended, for instance, under the name of its author, Randall Munroe. A major study about the sugar industry’s manipulation of health research trended under the unsexy acronym for the study’s journal.

On Sept. 20, an important analysis of Republican presidential nominee Donald Trump’s tax plan trended under the name “Tax Foundation,” a phrase almost guaranteed to repel mainstream interest. On the other side of the political spectrum, a report on Democratic nominee Hillary Clinton’s health trended not under the name “Hillary Clinton,” as is typical for political news in Trending, but under the pejorative Breitbartism #HackingHillary.

Crime stories also tend to get nonstandard names: Sometimes it’s the city where it happened; sometimes it’s a description of the crime itself. (This is the difference between “Tulsa police shooting” and “Tulsa, Okla.”) And the names of people involved in crimes virtually never trended in our sample. Instead of “Larnie Thomas,” we got “Edina, Minn.”

None of this is intentional, mind you. Like so much of Facebook Trending, the quirks and lapses of Topic names are the natural output of an automated system. Where Facebook previously had an editorial team tasked with rewriting topic names that were vague, inaccurate or unbalanced, power recently shifted to an algorithm.

That algorithm, which surfaces popular topics and stories for cursory editorial review — a process we discussed in greater detail last week — also identifies the word or phrase most often used by in Facebook posts about that story. Engineers can still intervene in certain cases, such as if the topic name is inaccurate, too long or poorly formatted. That said, the engineers have been known to miss things, and they only intervene in a minority of cases. Generally speaking, Facebook names its Topics the same way Twitter names its trends: If users are saying it, then it goes up verbatim.

Twitter Trends are not quite like Facebook Trending, however. Twitter makes no claim that its trends are based in “real-world events.” In fact, odds are that, at any given moment, one or more of Twitter’s 10 U.S. trending topics is a hashtag game or in-joke among 5 Seconds of Summer fans.

Among tech platforms with news aspirations more similar to Facebook’s, its approach to naming trends stands out. Apple, which identifies “Trending Stories” in its proprietary news reader and on the iOS 10 lock screen, defers to the exact headline of the featured article. Google, widely considered a leader in the space, relies on a sophisticated semantic network to identify topics adjacent to trending searches and stories, which it then clusters into series — as in “Shia LaBeouf, Mia Goth, wedding.” Google Trends has used that method since mid-2015, when it upgraded from a raw list of phrases that its users were Googling.

Facebook doesn’t claim to have cracked the Trending code yet. In fact, the site has been iterating the way it handles Topic naming since the departure of its editorial team in late August. First it swapped the team’s handcrafted descriptions for share metrics; then it tweaked the language that appears when you put your mouse over a trend. The next step will involve re-adding context to the Trending module … without hiring editors again.

“There are a range of approaches,” Will Cathcart, Facebook’s vice president of product management, said in a recent interview with The Washington Post. “We want to balance objectivity and the ability to scale, and we want to have something that users can scan that’s understandable. We’re making a lot of improvements around … cases where the system hasn’t given users what they wanted.”

As Trending continues to evolve, that may actually be its main challenge: How to get “the system” to more perfectly predict the complex expectations of humans.

Liked that? Try these!