In a study published Nov. 16 in Nature Computational Science, Northeastern University professor Babak Heydari and graduate student Hongan Zhu demonstrated how AI models can more accurately reproduce human decision-making. By factoring in human flaws and cognitive habits, researchers can build more realistic simulations to test ideas before consequences hit the real world.

Simulating Human Behavior Through Agent-Based Models

The concept of using models to simulate human choices traces back to a 1971 experiment by Nobel laureate Thomas Schelling. Using pocket change and dice on a grid to represent neighborhood residents, Schelling demonstrated how small individual choices—such as wanting neighbors from a similar background—gradually accumulate into large-scale urban segregation.

That early work laid the foundation for modern agent-based simulation. Researchers value these computational tools because conducting real-world human studies requires significant time and financial resources. Furthermore, simulations allow scientists and policymakers to evaluate high-risk scenarios without exposing actual people to harm.

"Human studies take time and money, which limits how many experiments we can run in practice," notes Jakob Kaiser, a senior researcher at the Nuremberg Institute for Market Decisions. Sangeet Khemlani, a senior cognitive scientist at the U.S. Naval Research Laboratory, adds that simulations bypass ethical concerns by testing scenarios that would be too risky to run on real populations.

Putting AI Agents to the Test in the Beer Game

To evaluate how well AI can mirror human judgment, Heydari and Zhu tasked agents from seven different AI models with playing the "beer game." In this classic simulation, four agents representing a retailer, a wholesaler, a distributor, and a factory manage a supply chain based on customer demand.

Each player operates with limited information, deciding how much inventory to order or supply while accounting for delivery delays. The goal is to maintain adequate stock without accumulating costly surplus or running out entirely.

Initially, left to their own devices, the AI agents made erratic choices, sometimes overcorrecting wildly even after demand leveled off. To bridge the gap, Zhu and Heydari introduced "cognitive grounding." This process involved supplying agents with mathematical models drawn from earlier experiments detailing how humans estimate demand, react to shortages, and handle shipments already in transit.

Across 231 test sessions, the cognitively grounded bots moved significantly closer to human benchmarks. In subsequent tests, the researchers even described psychological experiences—such as pressure, uncertainty, or the urge to overcorrect—rather than explicit decision rules. Agents from six of the seven models adapted and responded more like human participants as a result.

Practical Applications for Businesses and Policymakers

Accurate simulations of human judgment offer significant practical advantages for urban planning, commerce, and public safety. By testing interventions ahead of time, decision-makers can avoid costly errors or strategies that look promising on paper but fail in practice.

Potential real-world applications include:

  • Emergency Evacuations: Testing which instructions guide residents to safety versus those that inadvertently create traffic gridlocks.
  • Public Transit: Evaluating whether adjusted fare prices successfully encourage drivers to switch to buses.
  • Retail Strategy: Exploring whether relocating a pharmacy attracts new customers or cuts off existing clients.
  • Supply Chain Management: Anticipating bottlenecks and testing methods to prevent panic buying and inventory shortages.

Ultimately, aligning collective AI behavior more closely with human nature gives developers better tools to understand and guide these systems, helping researchers design practical improvements for complex real-world challenges.