Facebook defends hate speech moderation AI, claims 50% drop in 3 years

Social media giant Facebook hit back on critics, claiming that the prevalence of hate speech on the platform has dropped by 50 percent in the past three years. The company said that internal documents leaked to the media were “being used to create a narrative” that its automated technologies used to detect hate speech were inadequate, and that the company “deliberately” misrepresents its progress.

“This is not true. We don’t want to see hate on our platform, nor do our users or advertisers, and we are transparent about our work to remove it. What these documents demonstrate is that our integrity work is a multi-year journey. While we will never be perfect, our teams continually work to develop our systems, identify issues and build solutions,” Guy Rosen, Facebook’s Vice President of Integrity, wrote in a blog post.

The company’s post is a response to an October 17 article by the Wall Street Journal, which said that the company’s own engineers are doubtful about its artificial intelligence (AI) algorithms being effective against moderating hate speech. The article said that the company’s own employees estimated that it removes “only a sliver” of posts violating its rules.

Further, Rosen argued that only considering the amount of content removed as a metric for success was the “wrong way” to measure success. Instead, the company considers a metric called “prevalence” as the most important. Facebook explained this metric in a 2019 blog post in the following manner:

“This is similar to measuring the concentration of pollutants in the air we breathe. When measuring air quality, environmental regulators look to see what percent of air is Nitrogen Dioxide to determine how much is harmful to people. Prevalence is the internet’s equivalent — a measurement of what percent of times someone sees something that is harmful.”

Essentially, Facebook is arguing that while there may still be hate speech on the platform, it isn’t as visible anymore. Rosen’s post said that hate speech prevalence has dropped to 5 views per 10,000 on Facebook. He claimed that the company had gone from AI proactively detecting only 23.6% of the content it removed to over 97% today.

He admitted though that the proactive rate doesn’t tell the company what posts it’s missing, which critics have often argued is missing from the company’s metrics.