Warned alike, AI agents avoid the less-crowded road while people take it

Warned alike, AI agents avoid the less-crowded road while people take it is a study by Takahiro Ezaki, Naoto Imura, and Katsuhiro Nishinari published in arXiv:2609.30883 on September 25, 2026. The research demonstrates that adding a single sentence warning—stating that others might follow a routing tip—causes populations of 50 GPT agents to spontaneously coordinate on a single road, avoiding the nearly empty alternative.

This shared forecast creates a feedback loop where agents anticipate others will switch, leading to persistent concentration on one road. Consequently, average travel time rises from 64 to 95 minutes, a significant inefficiency where any individual agent could have saved 69 minutes by switching alone. This phenomenon, termed "collective inefficiency," persists across 100 rounds and remains robust even when explicit numerical feedback is provided.

In contrast, human participants respond differently to the same conditions. In all-human groups, participants maintain a near-balanced distribution of road usage. In mixed groups containing both AI agents and humans, people increasingly take the road that the AI agents avoid, effectively optimizing their own travel times while the agents bear the brunt of the congestion. The study highlights that shared information among homogeneous AI populations can sustain suboptimal outcomes, whereas humans adapt to counterbalance these AI-driven imbalances.

The paper examines how a minimal social cue changes collective decision-making in a congestion game involving GPT-based agents. In the experiment, agents receive routing advice and are also given a single-sentence warning that other agents are likely to follow similar advice. Rather than using that warning to anticipate congestion and choose the less-used alternative, the agents appear to treat the warning as a shared focal signal: when all agents receive the same social information, their choices become highly correlated, and the population converges on one road. The result is a substantial efficiency loss, with average travel time increasing from about 64 to 95 minutes.

A key insight is that adding social awareness to LLM agents can backfire. The warning is not simply neutral information about the environment; it changes the strategic landscape by making the other agents’ likely behavior more salient and common-knowledge-like. In this setting, the GPT agents’ behavior resembles herding or coordination failure: they avoid the less-crowded road even though that choice would improve aggregate outcomes, while the paper reports that people take the less-crowded road in the corresponding comparison. This suggests that LLM agents may respond to socially framed prompts through salience, imitation, or focal-point reasoning rather than by computing the welfare-maximizing or equilibrium-improving response.

The result matters because LLM agents are increasingly used in domains where decisions are made under social or systemic interdependence: traffic routing, scheduling, resource allocation, market forecasting, and multi-agent coordination. The study shows that prompt design can have collective consequences beyond individual task performance. A well-intentioned instruction to consider what others are doing may induce correlated choices, reduce diversity, and produce worse group outcomes than either private advice or no advice at all. This motivates caution in deploying socially aware AI agents and suggests that designers may need explicit load-balancing incentives, diversification mechanisms, or prompt structures that prevent shared cues from becoming self-fulfilling coordination signals.