It's a difficult question, frankly. There's no question that it remains a challenge.
The information on which the models are being trained—there's no question—have a wide assortment of biases embedded in them. There are zero questions about it. Also, the people building these tools are largely the same demographic that built the previous generation of tools, and we saw some of the implications of that. As well, the final outputs of these products are themselves showing signs of some of these biases. All three of those things are true.
The AI problem is slightly different, though, when you're referencing social media. This is not based on any technical expertise I have, but there is much more abstraction and distance between the design decisions made by programmers and engineers and the ultimate output of the technological system with AI than there was for social media.
With social media, there was a very close connection. You would tweak the algorithm for a newsfeed, and this would have a direct effect on how the newsfeed behaved. You could do a content moderation algorithm that had a direct effect on what the user saw and didn't see or what was prioritized and what wasn't.
With AI, there are two different layers of it.
One is a model-weighting issue, which people have some control over, but in a very opaque way. We are directing a system. We're not controlling it in the training of a model. It's not entirely clear what the effect of those biases would be in the training of the model.
The second is on the output of the model and the interface through which we consume it. This is probably where we need to focus more, particularly on audit attention. It's not necessarily auditing language models themselves, but auditing the outputs of the chatbots, which have a layer of design in them and are probably more subject to the kinds of biases you're talking about. I'd probably focus attention on that set of problems.
