Sarah Russell

It is a pleasure to serve with you in the Chair, Ms Butler. I thank the hon. Member for Dewsbury and Batley (Iqbal Mohamed) for securing this debate. There are two problems—maybe three—with AI. The first is that we do not distinguish very well between what is and is not AI. Although AI and tech are obviously related, they are not the same thing. It is important that when we talk about AI we distinguish it from tech. There is a need to regulate a lot of tech much better than we currently do, but AI poses very specific problems. The first one—I can see people from ControlAI in the Public Gallery—is the fact that we do not fully understand the models. It worries any sensible-thinking person that we are unleashing technologies that appear to be able to self-replicate and do other things, and we are incorporating them into military hardware without a full understanding of how they work. We do not have to be a catastrophist or conspiracy theorist to be worried. I am generally a very optimistic person, but it is important to be optimistic on the basis of understanding the technology that we use and then regulating it appropriately. That does not mean stifling innovation, but it does mean making sure we know what we are doing. When I look at AI, we have, as I said, two problems. One is rubbish in, rubbish out, and there is a lot of rubbish going into AI at the moment. We can see that in all sorts of terrible situations. We have a huge amount of in-built gender bias in our society. That means that, for instance, if we ask for AI to generate a picture of a female solicitor, as I am, we will get a picture of a woman who is barely clothed, but has a library of books behind her. That is not how female solicitors that I know go to work, but that is how AI thinks we are, and that has real-world impacts. If we ask AI to suggest an hourly rate as a freelancer, it is on average suggesting significantly lower rates for women than it is for men. There are questions about algorithmic bias permeating the whole of the algorithm. Questions have been raised recently about LinkedIn. I and a lot of women I know are finding that we have significantly less interaction via LinkedIn than we used to. Various women have now changed their gender on their bios to male and suddenly find that their engagement levels go straight back up. LinkedIn appears to think we are not interesting and people will not want to read our content, so it is stopping showing female content at the same rate, it would appear. I caveat that I have not been able to speak to LinkedIn directly, but certainly a lot of women I know are reporting these problems. We put in bio stuff to start with, but huge amounts of the image training data is based on what is publicly available on the internet, and that image training data of women on the internet is largely pornographic, which influences what comes out the other end of these models. When we look at that in terms of children, we have real problems. Nudification apps are huge and need to be dealt with. I would like to get into how I am worried about that and deal with health and how we do not have good enough training data on the interaction between gender and health and various other matters, but I will stop now. I thank everyone for their time today. I know colleagues will pick up important points.