Time is running out to get BOTH new Marketplace Sock designs for only $5/month.
Don’t wait – this special offer ends soon!
Esther Duflo is one of this year’s winners of the Nobel Prize in economics. She and her colleagues at the Massachusetts Institute of Technology’s Abdul Latif Jameel Poverty Action Lab and Harvard University won for their unique approach to tackling poverty. They try to break up huge problems like immunization or educational opportunity into smaller problems. Their team comes up with questions about what might work and then uses randomized controlled trials that are traditionally used in medicine and hard sciences to compare the impact of specific approaches. She said the team also uses artificial intelligence and machine learning to deploy the findings of their research in the most efficient way possible. The following is an edited transcript of our conversation.
Esther Duflo: [Machine Learning] gets learning to a deeper level. It gives you one more level of action, which is to say not just whether the program is effective on average, which you can get from the experiment, but where the program is the most effective, and therefore where a government with limited budget would want to expand it. It’s very useful, because otherwise the government will be saying, “Well, that’s nice. Incentives are effective, but they are expensive, so what am I going to do? How to choose where to put them?” So that’s where the machine learning really comes in.
Molly Wood: Do you worry about data integrity? We know that sometimes algorithms can reinforce bias. Do you worry that errors could be introduced by governments who are working with datasets that maybe just reinforce the old ways?
Duflo: I do worry about the bias that exists in the algorithm, [but] not for the moment because they are so underused in governments that I don’t think there is any damage that they are doing right now, but certainly for the future. The algorithms tend to reinforce existing biases, because they are learning from our previous experiences. One of the things that we need to be proactive about is to make sure it doesn’t happen. That’s where I think it’s essential to not think that the use of any artificial intelligence is going to be a substitute for policymaking and my full thinking, but it has to be a complement.
Wood: Is cost a barrier — cost of accessing this technology or creating these algorithms — for developing countries and their governments?
Duflo: I think it goes in the opposite direction. One of the big values of larger access to digital data processing, to the algorithm and to just a better way to handle the data in general, and to process the data the government has, will actually reduce costs and not increase them. I don’t think cost is the constraint. What is the constraint is more the amount of talent; we need people who are excited to work in this area and willing to join government to put their intelligence to the surface of this type of work. For example, when Nandan Nilekani from Infosys joined the government to start, in a sense, the crazy project of giving every single Indian person the unique identification number that was tied with biomarker identification. This might have sounded like a completely crazy plan, but because he had this intelligence and his wherewithal and the reputation, he could actually make it happen. And the infrastructure that has been built on the basis of the unique ID system is not without flaws and issues but has been, on balance, transformational for many of the ways that India is now delivering public services to its poor members.
Wood: Much of your work in developing these trials has been about, as you put it, finding the right question and asking questions that you can test in opposition to other ideas. Do you think that there is a time when some of this machine learning or these predictive tools could actually help you generate future questions?
Duflo: I don’t think so. I think what the predictive technology does is to tell you what is likely a good course of action given what you already know. I don’t think, so far, I’ve seen, at least, applications of using them to generate an entirely new set of questions. That, so far, has come more from the combination of field experience and relying on the mistakes made by others.
Wood: So far, we still need humans to do the thinking.
Duflo: Yeah, I think so. You need some amount of human intelligence for some time to come.
It’s notable that Duflo is the youngest person to win a Nobel Prize for economics at just 46, and she is only the second woman to be awarded the prize. But she’s very diligent in telling us and everyone else that her work is a team effort. She won alongside her husband and research partner, Abhijit Banerjee, and their longtime colleague Michael Kremer.
Over on the “Make Me Smart” podcast, we talked with her about how one of the things the work has done is upend what might be known as received wisdom. For example, in India and Kenya, they tested things like whether giving kids textbooks, feeding them meals at school, improving teacher quality, even smaller class sizes, would help students learn more. And interestingly, none of those things made a big difference. So they worked on making schoolwork more relevant to students’ actual lives and changing the way they hired teachers. And the program they recommended, according to the Nobel committee, is now helping improve the schooling of some 5 million Indian children.
And I must say, considering how when we talk about the tech industry, we’re very often talking about a small group of people with startup CEOs telling everyone else how things should be done, it’s kind of wonderful to find out that the best approach, actually, is to take that idea about what should work and test it against something else to make sure it works. You throw it out if it doesn’t work and go with the thing that is effective. I mean, that’s how this is all supposed to work, right?