Researchers tested shared forecasts in a two-road congestion game with GPT agents and human participants, finding that agents avoided the less-crowded road when warned about others' potential actions, leading to collective inefficiency and unequal burdens.