Skip to main content
Science Catalogs
Physical Sciences and Engineering

Teaching AI agents human decision flaws makes supply-chain simulations more realistic

AI agents built on large language models do not naturally make decisions the way people do, but they can be nudged closer. Babak Heydari and Hongan Zhu at Northeastern University report in Nature Computational Science that giving agents a description of how humans typically misjudge supply and demand brought their behaviour nearer to human benchmarks in a simulated supply chain.

Why simulating people matters

Shortages can feed on themselves. Panic buying of toilet paper in 2020 is a familiar example, and Heydari points out that everyone along a supply chain begins reacting to everyone else. Simulations that capture this behaviour could help planners anticipate disruption and test remedies before applying them in the real world.

The approach has a long history. In 1971 the economist Thomas Schelling used coins on a grid to show how modest individual preferences can produce large-scale residential segregation. That experiment helped lay the groundwork for agent-based modelling, in which simulated individuals stand in for real ones. Replacing human participants with software also avoids the cost, the delay and the ethical limits of studies on people, as outside researchers quoted in the Northeastern report noted. They also warned that a simulation is only useful if it is accurate.

The beer game test

The team used the beer game, a classic exercise in which a retailer, wholesaler, distributor and factory each decide how much to order or supply. Every player sees only part of the system, and deliveries arrive late, so each must keep track of stock in transit while avoiding both shortages and costly surplus.

Agents drawn from seven AI models played the game. Without guidance they made some peculiar choices, including continuing to over-order after customer demand had levelled off, where a human player would typically adjust.

Cognitive grounding

The researchers then added what they call cognitive grounding. Drawing on earlier experiments that analysed human play, they described how people estimate demand, react to shortages and excess stock, and sometimes fail to give enough weight to orders already on their way. Zhu stressed that this text did not tell agents how much to order; it was effectively a verbal version of a mathematical model of play.

Across 231 sessions, the grounded agents moved closer to the human benchmark, and some came close to matching it.

In a second set of tests, the team described what players might experience rather than how they decide: uncertainty, pressure to order more, or an urge to overcorrect. Agents from six of the seven models became more humanlike as a result. Some remained stubborn, however. One model cut its orders from 1,000 units to 100 or 50, a drastic swing of a kind that the grounding did not smooth out.

Limits and next steps

The evidence comes from a single stylised game, and the source gives no detail on how the results compare across models beyond the figures above. Whether the same technique transfers to richer settings, such as evacuation planning or transport choices, remains to be shown, although those are the applications the authors point to.

The work also has a bearing on AI safety and design. If the collective behaviour of agents can be made more humanlike through plain-language descriptions, developers may find it easier to understand and steer groups of agents. Zhu likened the method to psychology rather than neuroscience: the team observes behaviour and redirects it with language instead of examining a model's internal workings.

Heydari describes his lab as spanning engineering and computational social science, with the aim of turning better understanding into better-designed systems. The next test will be whether grounded agents stay faithful to human behaviour in settings where less is known about how people actually decide.

Source: Can teams of AI agents recreate the chain reaction of human decisions? (Tech Xplore, 22 November 2026).

Image: AI-generated illustration.

Working on this topic? Find a journal for your research or join the discussion on Scholars Post.