Altruistic AVs can reduce – or worsen – traffic congestion, new model shows

A new paper from researchers Lihui Yi and Ermin Wei (arXiv:2605.23782) tackles a pressing question for the future of urban mobility: how do autonomous vehicles (AVs) programmed to be altruistic affect traffic networks shared with selfish human drivers? The authors model the scenario as a simultaneous routing game where human-driven vehicles minimize their own travel time, while AVs act as altruists aiming to minimize the total social cost across all users. This is a departure from typical assumptions that all agents are self-interested.

Using a variational inequality (VI) framework, the researchers establish that an equilibrium exists even without convex cost functions, and that the aggregated link flow and social cost are unique for a specific class of cost functions. Crucially, they derive sufficient conditions under which increasing the number of altruistic AVs improves social cost, deteriorates it, or has no effect. This goes beyond previous worst-case bounds by pinning down exactly when each outcome occurs. The analysis also shows that a centralized social planner achieves the same equilibrium as the decentralized scenario under convex costs.

Numerical experiments illustrate how social cost changes with AV penetration rates and system parameters. The findings carry significant implications: while altruistic AVs can reduce congestion by taking longer routes to benefit everyone, they can also backfire if network topology or demand patterns create perverse incentives. The work provides a rigorous foundation for policymakers designing mixed-autonomy traffic systems and for engineers programming AV routing algorithms to avoid unintended consequences.