Ms Claire Rush:

I have personal experience of doing that content review work. I have been working in this area for many years. The role of content reviewers has really changed because of the evolving lead-in and the technological advances we have been able to make. Some eight or nine years ago all the content reviewed by content reviewers would come from human reports, that is, people clicking on a piece of content and indicating that they did not like it because they thought it was hate speech or violence. That would be reviewed manually. The volume of content has increased. Some forms of harm are particularly egregious and abhorrent. All types of harmful content hold their own in some ways, but it is easier to develop the technology for areas like terrorism and child sexual exploitation imagery, CSEI, because they are very visual and clearcut. Context does not really come into it. That is different from a piece of hate speech, for example. It is much harder to train classifiers or machine learning to tell if a hateful slur is being used in a derogatory way or in a comedic, satirical or self-referential way. That is why the rates of AI detection are lower for bullying and those types of content.

To return to the previous point, it is increasingly the case that content reviewers are not just reviewing reports from human users. The AI is always trying to learn and become more precise. Where it cannot take automatic action to remove content once it matches it, it will sometimes send content to humans for review as a double-check.