“People rarely make decisions in isolation,” he says. “They make decisions with algorithms providing them with suggestions.”
That goes beyond using Netflix’s recommender engine to choose a movie or accepting Amazon’s suggestion of products “you might also like.” Machine learning has now pervaded every profession. “Judges decide which defendants should await trial in jail based on predictions from algorithms; doctors decide which patients to treat using algorithmic recommendations,” Rambachan says. “Understanding decision-making requires us to take very seriously the fact that there are now algorithms in the loop based on data optimized in particular ways.”
Rambachan’s research examines how algorithms affect decision-making, and how they can be used along with human intelligence to make decisions better and fairer. In addition, he’s examined how algorithms can be used to improve behavioral research itself.
“We have a very long history in behavioral economics to know that people are very bad at making decisions,” he says. “We aren’t good at aggregating data and using data to form accurate predictions about what might happen.” On the other hand, he says, humans are good at making inferences from soft characteristics that aren’t always quantified in hard data. Algorithms, meanwhile, are very good at extracting signals from data sets to make decisions, even while they might miss more subtle qualitative information. “So the big question is, how can we get the best of both worlds?” Rambachan asks.
AI and human judgement combined lead to better judicial decisions
In one recent study, Rambachan applied that question to the criminal justice system, looking at how human and artificial intelligence might be used in concert to make difficult decisions about whether a defendant should be released pending trial. “That decision is supposed to hinge entirely on the judge’s prediction about whether this person will show up in court,” Rambachan says. The stakes are high, since incarcerating people unnecessarily can have devastating effects on their finances and mental health.
For the study, published in the Quarterly Journal of Economics in spring 2024, Rambachan and colleagues looked at more than 750,000 decisions made by judges in New York City. They found that in about 40% of decisions, judges made systematic mistakes, usually considering minority defendants, especially African Americans, to present more risk than they actually did, resulting in overly strict confinement. On the other hand, they found that using an algorithm by itself wasn’t necessarily the best strategy; rather, the most accurate results would come from using the algorithm to process defendants with characteristics where judges were likely to err, and then relying on human judgment for the rest. “That would actually improve results on both the status quo and algorithm only,” Rambachan says.
The same principles of algorithmic triage could also be applied to decisions made by doctors about what patients to treat and financial managers about which borrowers to lend money to, Rambachan says. That doesn’t mean such systems would be easy to implement in a real-world situation, however. “Judges, perhaps unsurprisingly, say, ‘I’ve been on the bench for 20 years, how can some data scientists tell me what is the right thing to do?’” Rambachan says. Still, he relishes the challenge of figuring out the best way to entice humans and machines to work together to achieve optimal results. “It’s just an amazingly exciting time to work at the intersection of economics and computer science,” he says. “There’s so much work to be done, and it has to be done interdisciplinarily.”
AI-driven insight into human decision-making
Another vein of Rambachan’s research addresses the use of algorithms in behavioral research, using AI to help economists better understand human decision-making. For example, a classic category of experiments in behavioral economics are so-called choice experiments, pioneered by Amos Tversky and Daniel Kahneman, in which humans are asked to choose between two options with uncertain results—for example, percentage chances of winning different amounts of money, or of saving different numbers of lives.
Depending on the choices and the way they are presented, economists have shown that human decisions are intuitive and averse to risk, rather than following the optimal rational economic models—a fact that can impact decision-making in finance, marketing, and policymaking. By training a machine-learning algorithm on such problems, Rambachan has found that AI has been able to derive novel scenarios for researchers to test, revealing new information about human behavior. “The way people are violating the axioms of ‘rational’ economic behavior in these algorithmically generated choices are different than the types of violations behavioral economists have documented before,” Rambachan says. Those findings, in turn, could lead to developing better methods of presenting data to people to achieve a desired outcome.
In all of these studies, Rambachan hopes that AI can enhance our understanding of human decision-making, allowing us to make better decisions than either humans or machines are capable of making alone. “We have a tremendous opportunity ahead of us, with a whole new suite of tools we can use to potentially improve the decisions made by judges, doctors, managers, and researchers,” he says. “My hope is that incorporating ideas from economics with tools from computer science can help us take advantage of this opportunity and accelerate the pace of progress in the social sciences.”