Ms Frances Haugen:

A simple example concerns even just disclosing which of the artificial intelligence systems exist today, and that covers most classifiers. This could be even just saying, “You must disclose which of these labelling systems exist [and hate speech, violent content and nudity are each a classification system] and which languages are supported, and give us a thousand examples of each score of those systems so that we can see how effective your systems are”. Just that one action about which safety systems, which languages and giving us some samples would allow us to see things. For example, Facebook's own documents talk about the idea that they treat all Arabic speakers as if they speak one language even though Arabic is actually six very different language families. There are things like the fact 75% of counter-terrorism content is classified as terrorism content. If we had something as simple as that disclosure on which safety systems and which languages and to show us some samples, we would be able to catch that. We would have been able to have a conversation saying things like, “Irish English is different enough to regular English and you do not support it as a dialect, but the Irish people deserve to be safe”. Those are the kind of things where ongoing disclosures would incentivise Facebook to invest in more languages or to invest in the quality of the safety systems in those languages.