AI removes 95% of offensive posts on Facebook

Facebook said the platform’s AI removes over 95% of posts that offend other users. However, the model does not yet understand memes.

Facebook has spent several years building an AI-powered algorithm that can quickly find and remove offensive posts on a social network. The company notes that technology can now detect 95% of the content without live moderators’ intervention. However, the remaining 5% is difficult to determine, and now the company is trying to figure out how to remove them.

Facebook said its AI system found 94.7% of posts (22.1 million posts) that contained hate speech, all of which were automatically deleted in Q3 2020. This is much more than before – in the same quarter a year ago, AI was able to find only 80.5% of posts (6.9 million publications) that contained hate speech. The company also published these figures in the Community Standards Enforcement Report.

Like many other social media platforms, we rely on AI to help moderators’ team make it easier to find inappropriate posts. It is a daunting, never-ending task – you need to remove unwanted custom messages and ads constantly. But the difficulty is that only people can distinguish a work of art from an erotic photograph and a subtle joke from obscenities. But soon, the AI ​​will be able to do it for the moderators.


The researchers note that the most challenging AI content is based on memes and clues that the model doesn’t define. Facebook is currently working on detecting hateful memes. In the spring of 2020, the company released a public dataset related to such content in the hopes of helping researchers improve their detection capabilities. As an example of content that may inconvenience other users, they showed a meme that depicts a cemetery with the text “here you belong.”

Artificial intelligence (AI) was unable to determine that this is offensive content. However, a few more years of training will allow the algorithm to understand why this humiliates some users.

Google News button