Facebook says AI is helping the platform remove hate speech and that its moderators are calling for more than 1,000 pages of content on anti-Semitism and racism to be removed from the site.
The German government is now considering a law that would impose a fine of up to 50 million euros on social networks such as Facebook if they do not remove hate speech and fake news quickly enough. Today, Utopia will present evidence to the subcommittee that Facebook could prevent the spread of hate speech. The world’s largest social network also agreed last summer to comply with European Union anti-bullying guidelines for its users and to require that racist comments be deleted within 24 hours. Facebook’s own terms of service prohibit all forms of bullying, harassment and threats. According to a New York Times report last year, German hate speech laws are much stricter than in other countries.
As a system to combat hate speech, Facebook has created a database of 10,000 memes that are part of its “Hateful Meme” database, a collection of memes that are supposed to be part of a campaign against hate speech. The data comes from an AI that is said to have detected more than 90 per cent of ‘hate speech content’ removed from Facebook. AI technology is being used to better detect and determine which memes violate Facebook’s policies.
On the platform, artificial intelligence is said to have proactively identified and stopped about 95 percent of hate speech, according to a Facebook statement.
That’s a comparison with a year ago, when Facebook proactively tracked only about 1.5 million hate comments. That’s less than 2.2 million in the first half of 2016, but still more than double the number reported by people. In other categories of “hate speech,” Facebook’s new use of automated systems has actually led to an increase in its overall rate of hate speech removal: 22.5 million articles were removed between April and June, up from 14.7 million in the same period last year, the company said. Facebook has removed about 4.1 million hate speech a month since early 2015, the year it began reporting hate speech, and deleted nearly a million more than at the end of last year. Although Facebook has removed almost twice as much hate-speech content as humans report, AI is still only able to capture about a third of all hate comments on the platform, or about 3.3 million a day.
This increase is due to an automated hate speech sniffer developed by Facebook’s artificial intelligence experts. This approach has helped the company address various forms of hate speech up to ten times faster than ever.
Facebook is already using AI to moderate its platform nudity to the point where users are challenging some decisions, but when it comes to hate speech, misinformation and other content, AI is still a work in progress. There is a huge amount of training data for Facebook’s AI systems, particularly in terms of context and nuance when it comes to picking up hate speech and even nuanced statements on platforms like Facebook. To pull something down, Facebook needs to teach A.I. systems what people see as hate, with all the contexts and nuances that go with it, Rosen said.
The company said it calculates the frequency of hate speech by selecting a selection of content to be seen on Facebook and showing how much of it violates the platform’s policies on hate speech. The company estimates that about 1% of what Facebook users see is about 0.5% of the content they see, and the company estimates the share of Facebook customers who have seen at least 1,000 hate comments in the past two years, according to Rosen and Rosen’s research team. It estimates that about 2 to 3%, or about 5 to 10%, of all content a Facebook user has seen is equivalent to 1 to 2 million to 4 to 5 million cases of hateful speech. They estimate that about 2-4% (or 10-15%) of this content has been viewed by Facebook, which equates to about 1.6 million or 2.7 million individual recorded pieces of hate speech.