In early December, reports emerged that Facebook is overhauling its algorithms used to detect hate speech. But now the social media conglomerate has explained how its new, advanced AI will differ from that of the past.
According to Facebook regional vice president Dan Neary, hate speech accounts for 0.1 to 0.11 per cent of all posts on Facebook—with the average user experiencing 10 to 11 views of hate speech that violate the platform’s content policies for every 10,000 views of content.
In addition, the detection rate for hate speech identified by Facebook’s AI before it is reported to the company is now at 94.7 per cent—up from 80.5 per cent a year ago and just 24 per cent in 2017.
The technology is, according to Neary, extremely effective at detecting what Facebook defines as hate speech. But recent reports from The Washington Post revealed Facebook has, for more than a year, been in the process to overhaul what it detects, and what it removes, as hate speech.
As The Post reports, this is because its algorithms were “race-blind”, with comments like “White people are stupid” treated the same as anti-Semitic or racist slurs prior to the overhaul.
The algorithms were not making a distinction between groups that were more likely to be targets of hate speech and those that have not been historically marginalised.
So, how will Facebook’s new methods of detection differ from those of the past?
“We know that hate speech targeted towards underrepresented groups can be the most harmful, which is why we have focused our technology on finding the hate speech that users and experts tell us is the most serious,” a Facebook spokesperson tells B&T.
“Over the past year, we’ve also updated our policies to catch more implicit hate speech, such as content depicting Blackface, stereotypes about Jewish people controlling the world, and banned holocaust denial.”
Thanks to significant investments in its technology Facebook now proactively detects 95 per cent of the content it removes.
The spokesperson added that it continues to improve how it enforces its rules as hate speech “evolves over time”.
“We have worked hard to understand the impact of our work on underrepresented groups as we know that hate speech targeted towards them can be the most harmful,” the spokesperson said.
The new overhaul, known as the ‘WoW Project’, is understood to have been running since 2019, as the result of working with external experts, external research and internal research, including frameworks like Susan Benesch’s Dangerous Speech project.
As a result, Facebook has made updates to focus its proactive detection technologies on catching the hate speech users and experts say is the most serious, including dehumanising speech, and expressions of contempt or inferiority.
However, these combined insights have seen Facebook stop using its proactive technology to find a small subset of attacks against White people, Americans, and men.
This is understood to have improved the precision of the hate speech Facebook does go after: the social media conglomerate now goes after more harmful hate speech, because its technology is more accurate at finding and removing this content.
The small subset of attacks that Facebook’s proactive technology is no-longer trained to find—phrases like ‘Men are Dumb’ and ‘Americans are Stupid’—will, however, still be removed if someone reports them.
This is understood to come in a bid to stop the perception of Facebook over enforcing on speech that is deemed less harmful, including people posting ‘Men are Trash’ after a breakup.