Rational Expectations, World Models, and the Future of AI

When the 2026 World Artificial Intelligence Conference (WAIC) convened in Shanghai from 17 to 20 July under the theme “AI Partnership for a Brighter Future”, it brought together more than 1,400 international guests across over 140 forums, alongside a High-Level Meeting on Global AI Governance. Among the voices was Professor Thomas J. Sargent, 2011 Nobel Laureate in Economic Sciences and Advisor to the Global Institute of Finance, Technology, and Society (GIFTS) at Nanyang Technological University, Singapore (NTU Singapore). In a keynote speech for the conference, Prof Sargent traced a direct intellectual line from the rational expectations revolution in economics to the world models that occupy today’s frontier of artificial intelligence research. 

World models sit at the centre of that frontier: systems that learn internal representations of an environment and use them to predict, plan, and act within it. Prof Sargent’s message to the AI community was that economics has been building and stress-testing world models of its own for half a century, and that each field now has a great deal to teach the other. 

What Economists Mean by Rational Expectations 
Prof Sargent opened with a disarmingly simple framing. Economics, he suggested, studies what economists can say about the economy, and economists use words in peculiar ways. The term he chose to examine was one he helped make famous. Rational expectations, he explained, is not a claim about human psychology but a technical device for constructing dynamic models that can be put on a computer. 

At the heart of the lecture was a distinction between merely descriptive statistical models, which Prof Sargent characterised as data compression devices, and interpretable structural models. A structural model describes artificially intelligent agents together with their environment, encompassing their preferences, the technologies they can access, their information sets, and the strategies of governments. Each agent inside such a model holds a subjective probability distribution over future outcomes, while the model as a whole induces an objective distribution. Equilibrium theorising, in his account, amounts to a mapping from everybody’s subjective beliefs to the objective behaviour of the economy, since how people think the economy works ultimately determines how it works. 

A Communism of Models 
The rational expectations hypothesis, which Prof Sargent attributed to the brilliant insight of John Muth, closes this loop by requiring that subjective and objective distributions coincide. The result is a massive fixed point equation and what he described as a communism of statistical models, in which every agent shares the model of the world that the model builder himself would use. He was careful to stress the “as if” character of the assumption, likening it to Newton’s account of planetary motion and to Milton Friedman’s celebrated description of an expert billiards player who behaves as if he had solved the underlying physics. The discipline of the assumption, he noted with characteristic wit, also renders economists immune to being asked why, if they are so smart, they are not rich, because the model builder forecasts no better than the agents inside the model. 

Where the Assumption Breaks Down 
Prof Sargent then turned to the limits of the framework, which is precisely where he sees the connection to modern machine learning. Rational expectations models assert consequences of events that could never be learned from data the model itself generates, an act of faith that becomes acute in the analysis of counterfactuals and historically unprecedented policies. Decades of work on least squares and machine learning in economies, much of it his own, have shown that adaptive agents do not generally converge to rational expectations equilibria. They converge instead to self-confirming equilibria, in which agents hold correct beliefs about events they observe infinitely often yet may remain mistaken about events rarely observed. He also described robust control theory, developed with his long-time collaborator Lars Peter Hansen, in which agents who do not fully trust their model surround it with an entropy ball and optimise against worst cases within it. 

Shared World Models and Machines That Discover Money 
For the AI researchers in the audience, Prof Sargent offered a striking reinterpretation. A modern AI world model, in his view, is a set of learned probability distributions describing an environment within which artificial agents apply tools for prediction, planning, and control. A rational expectations equilibrium, seen through that lens, is simply a shared world model. 

This convergence of fields, he reminded listeners, is hardly new to him. In the late 1980s at Stanford, he and his co-authors Ramon Marimon and Ellen McGrattan placed artificially intelligent agents built on John Holland’s genetic algorithm classifier systems inside a Kiyotaki–Wright search model of money, publishing the results in the Journal of Economic Dynamics and Control. The agents, trading bilaterally in an environment with no double coincidence of wants, discovered monetary equilibria on their own, in what he called an early multi-agent reinforcement learning experiment. 

Prof Sargent closed by describing new work completed this year with Ziyue Yang, in which the two revisit those decades-old questions using a tabular variant of the MuZero algorithm descended from AlphaGo. Do artificially intelligent agents learn to use something as money, and do they converge to rational expectations or settle into self-confirming equilibria? The exercise, he said, taught them something about both the limits and the potential of building world models. It was a lot of fun, he added, and it was not easy. 

Shaping the Future: Three Key Takeaways 
Three ideas from the keynote stand out for anyone building, deploying, or governing intelligent systems: 
Beliefs shape outcomes: In economies, how people, and increasingly machines, think the world works helps determine how it actually works. Rational expectations is the limiting case in which every agent shares a single world model, a discipline that keeps models honest even if reality rarely achieves it. 
Learning does not guarantee truth: Adaptive agents, human or artificial, tend to converge to self-confirming equilibria, holding correct beliefs only about what they observe often. Robust decisions must therefore withstand model error rather than optimise for a single trusted model. 
Economics and AI are converging: From genetic-algorithm traders in 1990 to MuZero agents today, placing artificially intelligent agents inside model economies reveals both what machines can learn about economies and what economists can learn from machines. 

These are also the questions GIFTS was founded to pursue. Prof Sargent serves as an Advisor to the Institute, supporting its mission to advance interdisciplinary research at the intersection of artificial intelligence, finance, and society. His keynote speaks directly to the Institute’s flagship Economic World Models Initiative: much as digital twins let engineers test physical systems before building them, the initiative constructs virtual economies in which firms, households, investors, regulators, and AI agents interact, disciplined digital laboratories where ideas like those in Prof Sargent’s lecture can be explored before decisions are put into practice.