At MIT's Laboratory for Information and Decision Systems, researchers have just proven something that nearly 100 years of preference prediction science got wrong: asking people to choose between three things, rather than two, unlocks hidden patterns in human desire that could transform how companies, governments, and platforms understand what we actually want.
The discovery centers on random utility models — mathematical frameworks descended from L. L. Thurstone's groundbreaking 1927 psychology paper — that have quietly powered decisions across industries for decades. These models estimate the "utility," or value, that people derive from choices, and use that information to predict behavior in what-if scenarios. Should a city allocate a $20 million budget to transit or schools? How will commuters respond if a major road closes for construction? What movies should Netflix recommend? Behind these questions lies a random utility model, trying to infer your preferences from your choices.
The conventional wisdom has always been to gather data through pairwise comparisons: Which do you prefer, Netflix or HBO? Coffee or tea? The reason is straightforward. "Assigning a precise numerical score, such as 4.37, to the benefit you get from a single item is very hard," explains Constantinos Daskalakis, the Avanessians Professor of Computer Science at MIT and a member of the Computer Science and Artificial Intelligence Laboratory. "Whereas comparing two things, and deciding which one you like better, is cognitively much easier to do." That logic has dominated preference prediction for nearly a century.
But Daskalakis and his collaborators — Gabriele Farina (an assistant professor in MIT's Department of Electrical Engineering and Computer Science), Yeshwanth Cherapanamjeri (now at Nanyang Technological University in Singapore), and Sobhan Mohammadpour (an MIT PhD student) — discovered a fundamental flaw in this assumption. When preferences are estimated from only pairwise choices, it becomes mathematically impossible to detect correlations between different preferences. Someone who favors gun control might also support government-sponsored child care; a fan of indie films might adore foreign cinema but dismiss Hollywood blockbusters. These connections matter enormously. If Netflix fails to recognize that correlation, it shows subscribers movies they don't want, risking lost subscriptions. A political campaign missing these links cannot target voters effectively.
The solution, revealed in a paper presented in April at the International Conference on Learning Representations in Rio de Janeiro, is deceptively simple: three-way comparisons. When large numbers of people rank three alternatives in order of preference, correlations between choices suddenly become visible. The same insights can also be drawn from a mix of best-of-three and best-of-two rankings. The mathematics proved what intuition might have suggested: human preferences are rarely isolated. They cluster and reinforce each other in predictable ways — but only if you ask the right questions.
The implications ripple across every industry that relies on preference data. Streaming services, e-commerce platforms, government agencies, and political campaigns all depend on these models to make consequential decisions. What makes this finding particularly striking is how obvious it seems in hindsight: of course preferences correlate; of course three choices reveal more than two. Yet that basic insight had escaped formal proof for nearly a century. As Daskalakis notes, when platforms remain "blind to the existence of such correlations," they cannot estimate preferences accurately — and the consequences, from poor recommendations to misallocated resources, cascade outward. The MIT team has now provided the key to seeing clearly.