Forecasting—making structured, probabilistic predictions about future events—is used in fields ranging from finance to public health to climate policy. Assigning likelihoods to outcomes, and updating them as new information arrives, can be useful anytime decision makers need to act under uncertainty.
At a session of the 2026 Behavioral Science and Policy Association Conference, held on the Harvard Kennedy School campus this June, scholars discussed how forecasting can be a useful tool for practitioners and what artificial intelligence means for the humans who have spent careers trying to get predictions right.
Philip Tetlock, a professor at the University of Pennsylvania who has researched forecasting for decades—including the forecasting tournaments that identified the best human predictors, known as “superforecasters”—spoke in a panel with Ezra Karger, director of research at the Forecasting Research Institute, who studies how well AI models forecast compared with human superforecasters.
They were joined by Juliette Kayyem, an HKS faculty member and former assistant secretary at the U.S. Department of Homeland Security, whose expertise spans crisis management, disaster response, and national security. Kayyem, the Belfer Senior Lecturer in International Security, approached the conversation from a practitioner’s vantage point: how predictive tools can reduce harm once disasters inevitably occur.
The panel was part of BSPA’s two-day conference, held at HKS this year and organized by conference co-chairs Professors Jennifer Lerner, Elizabeth Linos, and Julia Minson with support from the School’s Shorenstein Center on Media, Politics and Public Policy.
“I want predictive capability to take me up to the point of decision, but I want meaningful human judgment and human responsibility before deployment decisions are made.”
The conference theme “Strong Societies Need Good Information” was evident through the forecasting session: much of the discussion centered on how human overconfidence and the rise of AI complicate predictions that are used in policy decision making.
Human superforecasters
“We often have very limited predictive ability and then a great deal of confidence in the explanations we construct after the fact. There is a lot of hindsight bias and retrospective determinism at work,” Tetlock said.
Part of the problem, Tetlock argued, is language. Experts tend to reach for hedges like “a distinct possibility”—phrases that can mean anywhere from a 20% chance to an 80% chance and that let forecasters avoid being pinned down to a specific number. “For high-status forecasters, there is often very little incentive to make explicit numerical forecasts. The best they can do is tie; more often, they risk losing status if they are wrong,” Tetlock said.
Through research on forecasting tournaments, Tetlock identified what he calls superforecasters: a small group of exceptionally accurate predictors. What sets them apart isn’t insider knowledge, but habits of mind: they think in probabilities, they update their beliefs as new evidence comes in, and they’re willing to commit to specific numbers rather than hide behind vague language. Good forecasting, Tetlock said, rests on two properties: calibration and resolution.
The AI factor
Superforecasters are, in effect, the benchmark for human forecasting performance, which is why ForecastBench—a project that Karger leads and Tetlock advises—tracks AI systems against them to see how close artificial intelligence is getting to the best humans. “It is strange to spend so much time trying to improve human probability judgment when AI may soon be very close to, or beyond, the best humans,” Tetlock said.
Karger argued that the response shouldn’t be to abandon human forecasting but to reorient the field around usefulness. “Researchers often start with the literature,” Karger said. “That may be a poor starting point for policy relevance. Instead, we should start with questions that policymakers actually care about and ask how forecasting can help answer them.”
What happens in practice?
While Tetlock and Karger explained the academic underpinnings and concerns, Kayyem provided a practitioner’s perspective and concerns, asking what happens when a prediction meets a disaster.
Kayyem argued that disaster management has shifted from a “random and rare” mindset to what she calls a “fail safer” standard: success isn’t preventing bad events, which are often unavoidable, but minimizing harm when they occur. “The bad thing is going to happen. Success is measured by whether 100,000 people die or 4,000 people die. Four thousand deaths is not good; it is less bad. The delta—the harm avoided—is the standard of success,” Kayyem said.
She also drew explained categories of risk that officials confront: high-probability, high-consequence events, which is where officials spend most of their time; low-probability, low-consequence events; and low-probability, high-consequence events known as “black swans.”
On AI specifically, Kayyem identified the human control problem: predictive tools should inform decisions, but a human being needs to remain responsible before consequential decisions are carried out, just as is the case with, for example, autonomous weapons. “I want predictive capability to take me up to the point of decision, but I want meaningful human judgment and human responsibility before deployment decisions are made,” Kayyem said.
The future of forecasting—how policymakers can use it and what it means that AI may soon outperform the best humans superforecasters—was just one of the many sessions at 2026 Behavioral Science and Policy Association Conference, which also held TED-style book panels, a closing session on behavioral science for strengthening democracy with HKS Professors Danielle Allen and Nancy Gibbs, and more.
—
Photo by Mario Tama/Getty Images. Portrait by Martha Stewart.
More from HKS
Why Harvard Kennedy School is embracing disagreement
HKS experts teach and model how better disagreement can lead to better outcomes for everyone.
Featuring Julia Minson, Archon Fung
Twenty-five years after 9/11, the first battalion chief to send firefighters into the Twin Towers tells his story
The tragedy of 9/11 led Joseph Pfeifer HKSEE 2006, MC/MPA 2008 to learn and teach crisis leadership.
Lessons for leading across difference
Leading across difference takes courage, empathy, and skill. Throughout this issue of HKS Magazine, faculty, fellows, and alumni show what it takes to bring people together and drive real change. We capture some of their insights here.
Featuring Archon Fung, Julia Minson, Brian Mandell, Tarek Masoud, Nancy Gibbs