Facebook’s algorithms struggle to recognize hateful content

Facebook is not yet able to properly recognize hate-mongering content using algorithms. The social network removed 2.5 million posts of hateful content in the past quarter, but the majority only after manual checks.

Facebook admits this in the publication of its first transparency report. The social network says that of the 2.5 million hate speech messages, 38 percent were identified as such by artificial intelligence. The remaining 62 percent of posts were not addressed until users reported and moderators manually reviewed them.

“Our technique isn’t working very well for hateful content yet,” Facebook said in an announcement of the report. In addition to algorithms to recognize content that violates the rules, the social network employs 17,000 moderators, who decided in 62 percent of the cases that the messages in question were not acceptable.

Facebook claims that its algorithms are very efficient in other areas. For example, 837 million spam messages were deleted in the first quarter. According to Facebook, nearly 100 percent of these posts were found and marked as spam before users did. Facebook also removed 21 million posts with nudity or sexual activity. Of these, 96 percent were discovered by algorithms before being reported by users, according to the network.

The most important measure to fight spam is counteracting fake accounts. In the first quarter, Facebook deleted 583 million fake accounts, most of them minutes after registering. In addition, Facebook says it blocks millions of registration attempts every day. Facebook itself estimates that despite its efforts during the first quarter, 3 to 4 percent of all active accounts were fake accounts.

In total, Facebook deleted 866 million posts that did not meet the guidelines in the first quarter. By far the largest part is therefore spam. It is the first time that Facebook provides details about deleted messages. From now on, a report will be published every quarter.