Skip to main content

The Meta Oversight Board says Facebook’s automated image takedowns are broken

The Meta Oversight Board says Facebook’s automated image takedowns are broken


The board says a Colombian political cartoon highlights a weakness in the system

Share this story

A Facebook logo surrounded by blue dots and white squiggles.
The Facebook logo.
Illustration by Nick Barclay / The Verge

Meta’s Oversight Board says the company should be more careful with automated moderation tools, criticizing it for removing a cartoon depicting police violence in Colombia. The decision arrived as the board took on a set of new cases, including a question about a video of sexual assault in India.

The Oversight Board, a semi-independent body funded by Meta, considered a political cartoon depicting Colombian police officers beating a man with batons. The cartoon was at some point added to Meta’s Media Matching Service database, which meant Meta’s system flagged it automatically for takedown when users posted it. But when users saw their posts removed, they began appealing the decision — and winning. The Oversight Board says 215 people appealed the removal, and 98 percent of the appeals were successful. Meta, however, didn’t remove the cartoon from its database until the Oversight Board took up the case.

Automation can amplify the effects of one bad moderation call

That fact troubled the Oversight Board. “By using automated systems to remove content, Media Matching Service banks can amplify the impact of incorrect decisions by individual human reviewers,” the decision says. A more responsive system could have fixed the problem by triggering a review of the bank when individual posts featuring the image were successfully appealed. Otherwise, images banned based on one bad decision could remain secretly prohibited indefinitely, even when individual reviewers later reach a different conclusion.

It’s one of multiple Oversight Board cases questioning whether Facebook and Instagram’s automated moderation is calibrated to avoid overaggressive takedowns, and as in previous cases, the Oversight Board wants more methods of, well, oversight. “The board is particularly concerned that Meta does not measure the accuracy of Media Matching Service banks for specific content policies,” it notes. “Without this data, which is crucial for improving how these banks work, the company cannot tell whether this technology works more effectively for some community standards than others.”

It’s asking Meta to publish the error rates for content that’s mistakenly included in the matching bank. As usual for the board’s policy recommendations, Meta must respond to the suggestion, but it can choose whether to implement it.

Meta’s punishments for praising extremist groups are “unclear and severe”

The Oversight Board also addressed one of numerous incidents testing Facebook’s line between supporting extremist groups and reporting on them. It determined that Meta had erred in taking down an Urdu-language Facebook post reporting on the Taliban reopening schools and colleges for women and girls. The rule prohibits “praise” of groups like the Taliban, and the post was taken down. It was referred after an appeal to a special moderation queue but never actually reviewed — the Oversight Board notes that at the time, Facebook had fewer than 50 Urdu-speaking reviewers assigned to the queue.

The case, the board says, “may indicate a wider problem” with the rules about dangerous organizations. After several incidents, it says the policy appears unclear to both users and moderators, and punishments for breaking the rule are “unclear and severe.” It’s asking for a clearer and narrower definition of “praising” dangerous individuals and devoting more moderators to the review queue.

Meanwhile, the Oversight Board is seeking public comment on two cases: The first concerns a video of a mass shooting at a Nigerian church, which was banned for violating Meta’s “violent and graphic content” policy but may have had news value that should have justified keeping it up. Similarly, it’s interested in whether a video depicting sexual assault in India should be allowed to raise awareness of caste- and gender-based violence or whether its graphic depiction of non-consensual touching is too inherently harmful. The comment window for both cases will close on September 29th.