Repeat After Me: Humans Run the Internet, Not Algorithms
The saga of Facebook Trending Topics never seems to end—and it drives us nuts.
First, Gizmodo said that biased human curators hired by Facebook—not just automated algorithms—were deciding what news stories showed up as Trending Topics on the company’s social network, before sprucing them up with fresh headlines and descriptions. Then a US Senator demanded an explanation from Facebook because Gizmodo said those biased humans were suppressing conservative stories. So, eventually, Facebook jettisoned the human curators so that Trending Topics would be “more automated.” Then people complained that the more algorithmically driven system chose a fake story about Fox News anchor Megyn Kelly as a Trending Topic.
Don’t get us wrong. The Facebook Trending Topics deserve scrutiny. They’re a prominent source of news on a social network that serves over 1.7 billion people. But one important issue was lost among all the weird twists and turns—and the weird way the tech press covered those twists and turns. What everyone seems incapable of realizing is that everything on the Internet is run by a mix of automation and humanity. That’s just how things work. And here’s the key problem: prior to Gizmodo’s piece, Facebook seemed to imply that Trending Topics was just a transparent looking glass into what was most popular on the social network.
Yes, everything on the Internet is a mix of the human and inhuman. Automated algorithms play a very big role in some services, like, say, the Google Search Engine. But humans play a role in these services too. Humans whitelist and blacklist sites on the Google Search Engine. They make what you might think of as manual decisions, in part because today’s algorithms are so flawed. What’s more—and this is just stating what should be obvious—humans write the algorithms. That’s not insignificant. What it means is that algorithms carry human biases. They carry the biases of the people who write them and the companies those people work for. Algorithms drive the Google Search Engine, but the European Union is still investigating whether Google—meaning: the humans at Google—instilled this search engine with a bias in favor of other Google services and against competing services.
“We have to let go of the idea that there are no humans,” says Tartleton Gillespie, a principal researcher at Microsoft Research who focuses on the impact of social media on public discourse. That’s worth remembering when you think about the Facebook Trending Topics. Heck, it’s worth repeating over and over and over again.
Facebook’s ‘Crappy’ Algorithm
Jonathan Koren worked on the technology behind the Facebook Trending Topics. The bottom line, says the former Facebook engineer, is that the algorithm is “crappy.” As he puts it, this automated system “finds ‘lunch’ every day at noon.” That’s not the indictment you may think it is. The truth is that so many of today’s computer algorithms are crappy—though companies and coders are always working to improve them. And because they’re crappy, they need help from humans.
That’s why Facebook hired those news curators. “Identifying true news versus satire and outright fabrication is hard—something computers don’t do well,” Koren says. “If you want to ship a product today, you hire some curators and the problem goes away. Otherwise, you fund a research project that may or may not meet human equivalence, and you don’t have a product until it does.” This is a natural thing for Facebook or any other Internet company to do. For years, Facebook, Twitter, and other social networks used humans to remove or flag lewd and horrific content on their platforms.
So, Koren and about five or six other engineers ran a Trending Topics algorithm at Facebook headquarters in Menlo Park, California, and across the country in New York, news curators filtered and edited the algorithm’s output. According to Gizmodo, they also “injected” stories that in some cases weren’t trending at all. (A leaked document obtained by The Guardian, however, showed Facebook guidelines said a topic had to appear in at least one tool before it could be considered for the Trending module.) The setup made sense, though Koren says he privately thought that the humans involved were overqualified. “It always struck me as a waste to have people with real journalism degrees essentially surfing the web,” he says.
Trending versus ‘Trending’
When it looked like Gizmodo’s story was finally blowing over, Facebook got rid of its journalist news curators—then it promptly had to deal with the fake Megyn Kelly story. People blamed the more algorithmically driven system, but Facebook said all along that humans would still play a role—and they did. A human working for Facebook still approved the hoax topic over that weekend, something many people probably don’t realize. But they were outraged that Facebook’s review system, now without a single journalist employed, let a fake story slip through.
Koren says the whole thing was “a bit overblown.” And that’s an understatement. From where he was sitting, “there wasn’t someone within the company going ‘bwahaha’ and killing conservative news stories.” But even if there was an anti-conservative bias, this is the kind of thing that happens on any web service, whether it’s Google or Amazon or The New York Times or WIRED. That’s because humans are biased. And that means companies are biased too. Don’t buy the argument? Well, some people want fake stories about Megyn Kelly, just because they’re what everyone is talking about or just because they’re funny.
The issue is whether Facebook misrepresented Trending Topics. Prior to the Gizmodo article, a Facebook help page read: “Trending shows you topics that have recently become popular on Facebook. The topics you see are based on a number of factors including engagement, timeliness, Pages you’ve liked, and your location.” It didn’t mention curators or the possibility that the system allowed a story to be added manually. We could deconstruct the language on that help page. But that’s seems silly. Algorithms don’t exist in a vacuum. They require humans. Besides, Facebook has now changed the description. “Our team is responsible for reviewing trending topics to ensure that they reflect real world events,” it says.
What we will say is that Facebook—like everyone else—needs to be more aware of the realities at work here. Koren says that Facebook’s relationship to the broader issues behind Trending Topics was characterized by a kind of “benign obliviousness.” It was just focused on making its product better. The folks building the algorithm didn’t really talk to the curators in New York. Well, however benign its obliviousness may be, Facebook shouldn’t be oblivious. Given its power to influence our society, it should work to ensure that people understand how its services work and, indeed, that they understand how the Internet works.
What’s important here is getting the world to realize that human intervention is status quo on the Internet, and Facebook is responsible for the misconceptions that persist. But so is Google—especially Google. And so is the tech press. They’ve spent years feeding the notion that the Internet is entirely automated. Though it doesn’t operate that way, people want it to. When someone implies that it does, people are apt to believe that it does. “There’s a desire to treat algorithms as if they’re standalone technical objects, because they offer us this sense of finally not having to worry about human subjectivity, error, or personal bias—things we’ve worried about for years,” says Gillespie.
Sorry, folks, algorithms don’t give us that. Certainly, algorithms are getting better. With the rise of deep neural networks—artificially intelligent systems that learn tasks by analyzing vast amounts of data—humans are playing a smaller role in what algorithms ultimately deliver. But they still play a role. They build the neural networks. They decide what data the neural nets train on. They still decide when to whitelist and blacklist. Neural nets work alongside so many other services.
Besides, deep neural networks only work well in certain situations—at least today. They can recognize photos. They can identify spoken words. They help choose search results on Google. But they can’t run the entire Google search engine. And they can’t run the Trending Topics on Facebook. Like Google, Facebook is at the forefront of deep learning research. If it could off-load Trending Topics onto a neural net, it would.
But the bigger point is that even neural nets carry human biases. All algorithms do. Sure, you can build an algorithm that generates Trending Topics solely based on the traffic stories are getting. But then people would complain because it would turn up fake stories about Megyn Kelly. You have to filter the stream. And once you start to filter the stream, you make human judgments—whether humans are manually editing material or not. The tech press (including WIRED) is clamoring for Twitter to deal with harassment on its social network. If it does, it can use humans to intervene, build algorithms, or use a combination of both. But one thing is certain: those algorithms will carry bias. After all: What is harassment? There is no mathematical answer.
Like Twitter, Facebook is a powerful thing. It has a responsibility to think long and hard about what it shows and what it doesn’t show. It must answer to widespread public complaints about the choices it makes. It must be open and honest about how it makes these choices. But this humans versus algorithms debate is a bit ridiculous. “We’ll never get away from the bias question,” Gillespie says. “We just bury them inside of systems, but apply it much more widely, and at much bigger scale.”