How COVID-19 helped — and hurt — Facebook’s fight against bad content

When Facebook sent home its content moderators in March due to the COVID-19 pandemic, announcing it would rely on automation to at least temporarily do their job, the company predicted the decision would have a major impact on its ability to find and remove content that violates its policies. Now, according to newly released data from Facebook, we know just how big an impact it had.

During the second quarter of 2020 the company removed less than half of the child sexual abuse material from Instagram that it did the quarter before — not because there was less of it, but because the company was less equipped to catch it. And on both Facebook and Instagram, the amount of suicide and self-injury content it removed dropped precipitously too. On Instagram, it fell from 1.3 million pieces of suicide content removed last quarter to just 275,000 pieces this quarter.

But in other categories, like hate speech, Facebook's new reliance on automated systems actually led to a drastic increase in removals, from just 9.6 million pieces of hate speech removed from Facebook in the beginning of 2020 to 22.5 million pieces removed between April and June.

The drop in the removal of child sexual abuse material from Instagram wasn't due to a decrease in the amount of it on the platform. Neither was the decrease in takedowns of suicide related content. It was due to the limited number of human beings who were available to look at those posts, since, initially at least, they were all working from home. "The reason this content is challenging is because it's graphic content that, at home, is very hard for people to moderate," said Guy Rosen, Facebook's vice president of integrity. "We want to be very careful with the environment that people have in order to look at that content."

It's not that the human reviewers are required to spot all child sexual abuse material. Automated systems are already responsible for removing 97.5% of those types of posts that appear on Facebook. But according to Facebook spokesperson Emily Cain, human reviewers are critical when it comes to "banking" child sexual abuse material. That is, taking known images and logging them so that Facebook's AI systems can then go find and remove them.

"Without humans banking this content then our machines can't find it at scale," Cain said. "And this compounds after a while, so our content-actioned numbers decreased."

"Overall, this pandemic and this situation really reinforced to us that it is always people and technology working together," Rosen said on a call with reporters Tuesday. "We always need people who look and measure and help tune our automation to ensure that we're always up to speed and always up to date with how content is evolving."

The decrease in content removal is a blow to Facebook's ongoing efforts to fight the spread of child sexual abuse material on the platform at a time when the National Center for Missing and Exploited Children says that it's seeing an exponential increase in the number of reports about child exploitation. That said, the company did manage to remove more pieces of child sexual abuse material from the Facebook app than it did last quarter. And yet, overall, in 2020, removals in that category are down significantly from where they were at the end of last year.

During the COVID-19 crisis, Rosen said Facebook has developed a ranking system to prioritize the most critical content in these sensitive categories. That might include anything from a live video to a post in which someone indicates they plan to harm themselves imminently. This ranking system was already in the works before COVID-19, but Rosen said the company expedited its development in response to the crisis.

"This enables our teams to spend their time on the cases where we need their expertise the most, and it means there will be a shift towards more content being initially actioned by our automated systems," Rosen said.

As for the sharp increase in the amount of hate speech being removed from the platform, Rosen attributed that, too, to the ongoing development of Facebook's AI systems. Because hate speech is less graphic than, say, a video of child abuse, moderators are more able to handle that content remotely. As Facebook's chief technology officer Mike Schroepfer told Protocol in a tweet, "The more … sensitive and nuanced the content the more we need help from people."

Of course, the perennial question about hate speech, child sexual abuse material and other types of problematic content is not just how much Facebook is taking down and how fast, but how prevalent that content is to begin with. On the subject of hate speech, that's a question that Facebook hasn't been able to answer yet. Turns out, measuring prevalence requires a lot of human input, too.