A few years back, several studies from political science researchers found that machine learning seemed to be able to do something miraculous: predict civil wars. The study results apparently showed that new AI models outperformed older statistical methods for figuring out which countries or regions were most likely to have conflicts break out.
But when Sayash Kapoor at Princeton went to reproduce these studies with a colleague, the results didn’t stand up.
AI models are supposed to learn from one batch of data, then be tested on how well they make predictions about another batch they’ve never seen before. But Kapoor found these evaluation protocols had not been followed. In several of the studies he looked at, the AI had been accidentally exposed to information that should have been hidden until the final test. It’s a bit like grading a student after they’ve already seen the answer.
When accounting for this methodological flaw, Kapoor discovered that the new civil war prediction models weren’t any better than the decades-old statistical models. Suspecting that political science researchers weren’t alone in these mistakes, he conducted a wider review, ultimately finding nearly 300 papers across fields from psychiatry to nutrition that contained similar flaws.
One medical study that claimed AI could predict high blood pressure, for example, turned out to be feeding the model information about whether patients were already taking blood pressure medication, effectively giving away the answer.
Thus began Kapoor’s role checking the AI industry’s hype against the evidence.
Kapoor grew up in Patna, a city in northeastern India, and studied machine learning at university. It was all quite theoretical until he jumped into a job at Facebook building AI models that would predict if children were at risk of abuse or suicide. “Pretty heavy stuff,” he says. Seeing AI’s potential impact on the real world led him to a PhD at Princeton’s Center for Information Technology Policy, which he finished in July 2026. In August 2027, when he’ll still be just 30, he’ll join the faculty of UC Berkeley to research AI’s impact on society.
In an era of sweeping claims about what AI can do, Kapoor checks the receipts. When the field of “predictive optimization” promised that AI could determine whether a criminal defendant should be released before trial, for example, Kapoor and his colleagues outlined specific flaws in the technology that its proponents would have to account for. And when AI executives promised that AI would make human work obsolete, he helped chart how AI’s impact on society is more likely to be akin to other “normal technologies” like electricity or the internet.
In this way, you might call him a dissenting voice. But given the AI industry has no shortage of evangelists, Kapoor might just be leveling the scales.