Deena Mousa - What AI Means for the Developing World

So much of AI discourse is centered on rich countries, their workers, their firms, their labor markets. But what happens when you shift the lens to less developed countries with different needs and access to AI? My guest today is Dina Musa, program officer at Coefficient Giving and one of the sharpest writers on how AI intersects with

global development. Here's our talk.

Welcome to The Economics of Work. I'm excited to welcome Dina Musa. Dina is a program officer at Co-Fishing Giving, formerly Open Philanthropy, and is one of the deepest thinkers and writers out there about how AI affects development. Dina, welcome to the show.

Thanks so much for having me, Ben.

So I want to start with a question that I think has got a lot of questions underneath it, which is why are there so many radiologists in the world right now?

It's a great question, Ben. And I think one that a lot of people are wondering, especially given it seems like one of those professions that is the most automatable, if any, of the white collar work. So it is completely digital. There's not a lot of pen and paper involved in radiology anymore. It's very based on pattern recognition, which is something we think of as AI being very good at. And it's mostly done, we think, like in isolation, basically. If you read a scan, you don't really have to be collaborating with a lot of different people. So it seems like almost like the perfect use case for AI automation. But I think it turns out for radiology, as for I think we'll find soon most professions, not so simple. I think there's quite a big gap between benchmark performance and actual clinical performance in ways that are underrated. So you can have a benchmark, and we have, that have said models are significantly better at humans at detecting a particular finding on a particular scan. But that doesn't necessarily translate to replacing their work entirely. So one of those reasons is they're often trained on a particular sample data set. So they're trained on a set of like one or maybe if you're lucky, a couple of hospitals. And that means that they generalize poorly when they're brought to new locations. There is a particularly funny example where they found that a model was recognizing when chest scans were done with a portable scanner versus done in the radiology suite because the images look slightly different to one another and as you can imagine often sicker patients are given the portable scanner and so it was generalizing from that that if you're using a portable scanner you're much more likely to have pneumonia and i think that sort of demonstrates a little bit of this black box of ai we don't really know how it's coming to the decisions it is and so this makes it much harder to generalize in new contexts or when you have new demographics in a particular hospital

Yeah, it kind of reminds me of the early days of Kaggle competitions when the winner would be someone who hacked the user ID of something to find out that the ID number was that they were really in one class versus another class. I wonder if the fact that it doesn't generalize well is actually really just a problem with classified machine learning as a tool. So when