Social media has always struggled with the question of how to deal with hate speech, but a new Online Hate Index launched by the Anti-Defamation League (ADL) this week showcased the full extent of the failure to confront it.
According to the ADL, this machine-learning system combs hate targeting marginalized groups on social media platforms, representing the first independent, cross-platform metric of antisemitic content therein. In its first version, the system discovered that two such platforms, Reddit and Twitter, failed to remove more than 70 percent of antisemitic content ADL detected on their platforms nearly a month on.
“For the first time, we are using the combined powers of artificial intelligence and ADL’s expertise to uncover antisemitic content online at scale,” Jonathan Greenblatt, ADL CEO, said. “Technology companies are not sufficiently transparent about the effectiveness of their anti-hate policies and product interventions. So we created a tool that offers a clear picture of the state of hate on social media in real-time, and we will use this tool to hold those social media platforms accountable for how well they proactively take down hate and how well their content moderators respond to reports.”
However, Greenblatt also noted that both companies have made substantial strides in addressing hate online, but they have more to do. Instead, he pitched the Online Hate Index as fuel for recommendations to help them tackle the broader societal problems stoked by hate and antisemitism. Reddit and Twitter were chosen, in part, because they are more transparent than most social media platforms.
As to the Online Hate Index, the system is trained over time to identify antisemitic language. It uses connections between social media content and language that human guides label antisemitic. Repeated interaction with labeled data trains the model to become more accurate, eventually enabling it to analyze English-language text and predict whether it includes hate speech far faster and over a far larger content volume than human moderators could match.
Complicating matters for traditional, human moderators is that social media companies themselves are the ones who typically report hate speech. They tend to rely on internal metrics the ADL described as neither verifiable nor useful in comparing platforms. Further, tech companies do not tend to provide data on who is targeted by identified speech nor identify the groups to which they belong. ADL pointed the finger at Facebook in particular for this, noting that the company tends to hide what it does know about who is targeted by hate speech, even from its auditing teams.