Study: AI agent trained on explicit social norms coordinates with humans nearly 4 times better

In July 2026, researchers showed, using pedestrian-driver interactions, that formalizing unwritten social norms and building them into an AI agent improves its coordination with a human in dynamic situations by almost four times compared with a baseline policy.

Why AI struggles to get along with people in the moment

People constantly coordinate with one another through implicit social norms that are hard to formalize: shared, unspoken expectations among the participants in an interaction. As AI agents, including large language models, become more embedded in everyday life, they also take part in such dynamic interactions, but they often coordinate with humans inefficiently, inattentively and "unnaturally."

The authors suggest the cause lies in how such models are usually trained: their behavior is fitted to human demonstrations, but the norms that produce that behavior are not explicitly isolated. In other words, the model copies the outcome but does not understand the rule behind it.

How the hypothesis was tested

To test it, the researchers chose pedestrian-driver interaction, a typical example of dynamic coordination in which both sides constantly read each other's intentions without words, and built a simplified experimental platform that reproduces the key features of such interaction.

• They collected 3,456 dynamic human-human interactions through this platform
• From these data they identified three principles underlying social norms: outcome predictability, value alignment and advantage awareness
• These three principles were explicitly built into the AI agent's behavior
• The effect was tested in a closed-loop task, that is, in real dynamic interaction with live people, not in a static test on historical data

How much coordination improved

In the closed-loop task with humans, the AI agent trained with social norms scored almost four times more total points than the baseline policy that ignored norms. Moreover, this agent outperformed even human-human pairs in coordination quality, by 43%.

What this means

The results suggest that turning unwritten social norms into explicit, measurable principles, rather than merely copying behavior from demonstrations, may become a workable way to make AI agents more natural partners for people in dynamic interactions: from autonomous vehicles that must "negotiate" with pedestrians without words, to any other scenario in which robots and assistants have to coordinate with people in real time.

An important detail: the authors tested the effect not on historical recordings but in a closed loop. The agent actually interacted with live people in real time, rather than just predicting what a person would have done in a recording of a past interaction. This sets the result apart from many laboratory tests of human-machine interaction, where a model is judged by how well it matches already recorded human behavior, not by the quality of coordination in real time.

Frequently asked questions

How many interactions were collected for the study?

The authors collected 3,456 dynamic pedestrian-driver interactions through a purpose-built experimental platform.

Which three principles underlie the social norms?

The researchers identified outcome predictability, value alignment and advantage awareness as the key principles that explain human social norms of coordination.

How much better did the AI agent coordinate after being trained on norms?

The agent that accounted for social norms scored almost four times more points than the baseline policy and outperformed coordination between humans by 43%.