As part of an event at Lemuria Books (Jackson, Mississippi), I had the chance to sit down with Mississippi Today to talk about the gender data gap, the reaction to Invisible Women and what comes next.
You can read an excerpt from my interview below:
“MT: You talk about how this is an age-old problem – we live in a world made by men with men in mind. Can you tell us why, in a world that increasingly relies on “Big Data,” it matters so much more? How it becomes deadly, even?
CP: Yeah, so I mean, the gap in data for women is already deadly, if you’re thinking anywhere from car design to health care, but the real danger is becoming exponential, because of the introduction of AI into every single part of our world. And the problem with developing AI using bad data, biased data, is that machine learning is not like a human, in that it doesn’t simply reflect our biases back at us – it amplifies them.
I’ve read so many papers since “Invisible Women” came out where researchers will be like, “we’ve developed this AI and it performs better than a radiologist at detecting lung cancer” or “can predict heart attacks five years before they happen,” and then when you look at the paper, not only are the datasets incredibly male-biased, so you’ve got that bias already baked in, but also, they’re not even thinking about sex.
One paper I’m thinking about that came out shortly after “Invisible Women” was published was about predicting heart attacks. And there are sex-specific risk factors. So, if you’re going to be predicting heart attacks in men versus women, you don’t want to have, as this paper did, something like a 70% male dataset, but you even more don’t want to have that data all mixed up together. Because that’s not going to work for men or women. And yet, there was absolutely no mention of sex in the paper. So, that is frightening. Because the problem with that is it could make the situation worse.
When I find AI exciting is when researchers are using AI to address problems that we aren’t addressing otherwise. So, for example, one woman I spoke to was developing AI to detect victims of domestic violence via injury patterns, potentially years in advance of them ultimately having to be taken to a shelter or something. Because of course victims don’t necessarily report, and it’s not something that we’re investing a lot of money in in health care – because there’s not a lot of money in it and doctors don’t necessarily have the time to do the sort of questioning of a victim, et cetera. So there is exciting potential for AI. But if we’re just using it to do what we’re already doing but faster, that’s where the massive pitfalls are.”