AI agents built on a few shared models serve many users. A shared forecast about others’ actions can align individual choices and reshape the allocation of scarce capacity. We examined this feedback loop in a two‑road congestion game.
In the first experiment, a population of 50 GPT agents received a one‑sentence warning: “Others might follow this routing tip.” All agents crowded onto the same road while the alternative remained nearly empty. The average travel time rose from 64 to 95 minutes, even though any single crowded‑road agent could have saved 69 minutes by switching alone. The warning thus discouraged the very move it predicted, and the pattern persisted for 100 rounds. Two other model families showed the same shift, though they did not lock onto a single road.
In twelve all‑human groups (240 participants), numerical reports or the tip and warning kept the traffic near balance. In twenty‑four mixed groups (another 240 participants plus agents), imbalance grew with the share of agents, while humans increasingly took the road the agents avoided.
Collective costs stayed below the all‑agent reference, but with a configuration of 15 agents and 5 humans, agents averaged 80 minutes versus 44 minutes for humans, revealing an unequal burden. Shared forecasts can therefore sustain collective inefficiency among similar agents, and a better group average can mask who bears the cost. Evaluations of AI agents that share resources should test population‑level dynamics, treat messages as interventions, and report cost distribution.
Review