The power of collaboration: How we can reduce traffic congestion

Details

Published in Nature Cities, this paper from Google Research tests whether an AI routing system designed for individual trip efficiency can be adapted to optimize city-level traffic flow. The core tension: individually optimal routing tends to funnel drivers onto the same fast arterials, worsening congestion for everyone. The experiment ran a switchback design across 10 US cities over six months, periodically switching Google Maps between baseline routing and coordination-aware routing that spreads traffic across the network. Only a small fraction of drivers are rerouted — enough to shift aggregate flow without imposing major detours on individuals. Results show that coordinating even a small portion of trips reduces congestion measurably for the entire network, including non-Maps users. The effect persists across cities with different street topologies. The key insight is that network-aware routing produces a cooperative equilibrium — unlike Nash equilibrium routing where each agent acts selfishly — and the benefits extend to people who are not part of the coordination system. This is a rare real-world causal demonstration of cooperative AI acting as a positive externality at city scale.