Beyond the Brainpower

Artificial intelligence’s real test is not whether a machine can outthink a person. It is whether the systems built around that machine help people make better decisions. At the 2026 INFORMS Annual Meeting, Michael Jordan urged decision scientists to shift attention from ever-smarter software toward the incentives, institutions, and relationships that determine how technology affects human outcomes.

For Jordan, a researcher at Inria Paris and Professor Emeritus at the University of California, Berkeley, the starting point is an engineering question: What problem are we trying to solve? Concrete goals such as more efficient transportation or better medical outcomes offer a more useful measure of progress than arguments about whether a machine possesses intelligence.

Large language models belong in that picture as channels to collective human knowledge. Their answers draw on the work and experience of many people. More capable channels can be valuable, but capability alone does not ensure that the larger system serves its participants well.

“The LLM is a channel, and it's a smart channel,” Jordan explained. “But the intelligence is often really coming from the collective.”

From Smart Machines to Intelligent Systems

The missing foundation, Jordan argued, is a partnership among statistics, economics, and computer science/optimization. Academia has developed strong connections between pairs of these disciplines through machine learning, econometrics, and algorithmic game theory. Real-world problems, however, sit at the intersection of all three. Algorithms must contend with uncertainty and with participants whose interests may not align.

“More capable channels can be valuable, but capability alone does not ensure that the larger system serves its participants well.”

That becomes especially important as software agents act for people. Those agents inherit human goals, encounter unequal information, and receive data from sources of varying credibility. Neither strategic behavior nor statistical bias disappears as software improves. Designing a useful agent therefore requires more than specifying what it can do; it requires understanding why participants will cooperate and when evidence deserves trust.

The challenge is to shape interactions, not simply improve an isolated prediction or optimize one participant’s decision.

A Contract for Better Decisions

Statistical contract theory offers an early framework for that task, said Jordan. By giving participants a menu of terms connecting data collection, costs, and rewards, a system can make private knowledge useful without pretending to know everything about each participant. The aim is to unite economic incentives with reliable statistical decisions, rather than treating those objectives as separate problems.

The broader agenda also reaches policymakers. Regulation can shape the conditions under which participants interact, steering systems toward better outcomes without dictating every individual choice. Jordan called on mathematical researchers to engage with government and help translate these relationships into workable rules.

For the INFORMS community, said Jordan, the invitation is to expand the definition of progress. Optimization remains essential, but so do equilibrium, inference, and mechanism design. AI’s consequential advances may depend less on what an individual agent knows than on whether a system makes honest participation worthwhile. Better knowledge matters most when the surrounding incentives help turn it into better outcomes.