A Quick Guide to Ethical AI with Dr Kerry McInerney

So you mentioned bad science, bad data, bad deployment.

Can you briefly explain those three problems?

Yeah, so I included in my talk and didn't really have time to elaborate on a range of other ways beyond proxies, all these pieces of data that kind of relate to sensitive personal characteristics, and historic data, all this problem of existing societal divides is directly replicated in training data sets.

I think those are the two that have historically gotten a lot of attention.

But I also pointed out a few subtler ways that we can see inequality being exacerbated potentially by poorly designed, poorly made AI systems.

And yeah, the first is just bad science, right? So this was particularly, I think, overt in say around 2020 when there was a lot of research into facial recognition.

We started to see grossly overstated claims about what these systems could do.

And some of them were just really poor science or replicated pseudoscientific ideas from the past, a kind of fake science that we've really debunked and thrown in the bin.

So one of these fields of fake science that we saw getting a little bit of a rebrand and a resurgence through AI was physiognomy.

So this is the study of And it even sounds very garbage when I say it, but looking at someone's face and using that to infer something about their personality or their personal characteristics.

So if you were to do that for me, you might say, oh, well, I think from the shape of your eyebrows and your nose that you're a very studious person, something like that.

They're entirely spurious, right? This was a very popular kind of what we call race science in the years gone by.

And so when we say race science, we mean these kinds of pseudoscientific fields that were used to justify forms of racial oppression.

And it was also very tied and had a resurgence around eugenics as well in the early 20th century.

So these are really bad fields.

We don't want to be giving them any more credence.

But come 2020, we start to see facial recognition projects that do exactly that.

They say, I've designed an algorithm that can deduce whether or not someone is a criminal from their face, to which most of us would say, what? That has literally no correlation.

But the problem is that if an algorithm does it, people might think that there's more truth to that story.

They'd say, oh, well, if a machine can find this correlation, then surely it must be true.

And we start to see a lot of different examples of this from, you know, a team at Stanford creating an algorithm that could theoretically determine whether or not you were straight or gay from your face.

You know, through to a lot of tools that claim to be able to deduce your personality type or style based on your face and its appearance.

So this wasn't just a kind of one drop in a bigger ocean.

This was becoming more of a trend.

And so, you know, for me, it's not just that we have these objectively bad tools that really can't do what they say they can do, but that they might reanimate old ideas that we've said are actually quite dangerous.