Ms Frances Haugen:

This is a really important question, and I am so glad that the Senator brought it up. We spoke before about the idea that because Facebook operates behind a curtain, it gets to set the rules of conversation. The ways we are talking about safety today have been shaped by Facebook's PR machine. They have put the focus on the idea that AI and censorship, or the removal of content, is the solution. They have said that the arguments we should be having are not about whether Facebook should be able to operate in the dark but about whether we are taking down too much content, or if we need to take down more. They know that when we argue about censorship, we will never come to agreement. However, we could have a conversation about the rules of the road, for example about whether Facebook should be allowed to operate without us being able to ask questions, but they do not want to have that conversation. Therefore, when Facebook comes out and talks about AI as the solution, they are saying that they can come in and write a magical hate speech classifier or violence inciting classifier. A classifier is a machine learning system or an artificial intelligence system. It allows you to say that a post is inciting violence, so it must be taken down. Facebook's own research says this strategy is deeply limited because language is nuanced. As the Senator said, a sentence that could be totally innocuous in one part of a country could be very different and defamatory in a different part of it.

Even for American English - let alone Irish English, British English or Indian English - it can only take down 3% to 5% of hate speech. Facebook's own documents said that because of these nuances in language, in a best-case scenario it would only be able to take down 10% to 20% of hate speech. This is for the languages in which Facebook has chosen to invest. What they did not tell the world was that in order to have artificial intelligence keep us safe, you have to rewrite those systems over and over again in each individual language. Think of how many member states exist in Europe. Think about how many languages exist in Europe. Think of how many dialects exist in Europe. A dialect has to be retrained. Facebook has chosen not to invest in all the languages of Europe. It has definitely not chosen to invest in all the dialects in Europe.

In 2018, Mark Zuckerberg wrote about engagement-based ranking, which is a kind of ranking whereby Facebook sees if it can get you to react more and rewards content that gets you to react more. He said that is dangerous because people are subconsciously drawn to extreme content. Even when Facebook asks people afterwards if they enjoyed this content, they say they did not. Subconsciously, our brains are evolved to be drawn to that content. This is super dangerous. It is disrespectful to a linguistically diverse world. Instead, we need to focus on how the algorithms work because those changes can apply to every single language and dialect separately. What was the Senator’s next question?