Efficiency and Equilibrium in Network Games: An Experiment

GAICIL Insights

The same people, playing the same game, cooperate beautifully in one room and abandon cooperation in the next.¹ The only thing that changes is who is connected to whom. Efficiency and Equilibrium in Network Games: An Experiment, by Edoardo Gallo and Chang Yan, studies a network game with a unique Nash equilibrium in which agents can earn higher payoffs by following a collaborative norm, and asks how network structure decides whether the norm survives.

The results draw a clean gradient. In the circle, a simple, symmetric network, subjects establish and maintain the collaborative norm. Add one hub connected to everyone, the wheel, and the norm weakens. In complex, asymmetric networks of 15 and 21 nodes, the norm disappears and play converges to the Nash equilibrium. Given the richness of the networks and the size of the strategy space, the authors present the convergence as strong validation of the theory of Ballester and coauthors, and as evidence that network position can be a key determinant of equilibrium play.

A second finding concerns perception. Subjects base decisions on their degree, their number of direct connections, more than on overall network structure. In the 21-node network, among nodes whose local features are not highly correlated with their overall position, degree significantly predicts play, while Bonacich centrality, which determines equilibrium play, does not. People navigate by the streetlights they can see.

AI enters on two fronts. First, digital markets are filling with networks of human and machine agents: recommendation systems, trading algorithms, autonomous agents negotiating with one another. The experiment shows that cooperation sustainable in simple, symmetric structures can collapse as networks grow complex and asymmetric. As AI agents multiply, the shape of their interactions will decide whether efficient collaborative outcomes emerge or equilibrium behavior takes over. Designers of agent ecosystems are, in effect, choosing between the circle and the 21-node tangle.

Second, the local-information finding hints at a role for AI as a navigator. People in complex networks focus on immediate connections because processing global structure is hard. AI tools can compute global properties such as centrality that humans struggle to perceive, helping participants grasp their position and the consequences of their actions. Whether such information would sustain collaborative norms or speed convergence to equilibrium is an open empirical question that experiments like this one can answer. A map can help travelers cooperate, or help them defect more efficiently.

The authors note the results align with public goods experiments showing that asymmetry across subjects lowers contributions. Asymmetry erodes cooperation. That resonates with debates over platform and AI market structure, where hubs connect to many participants. Hubs change the strategic environment for everyone, weakening norms that symmetric structures sustain. Platform design, including the design of AI intermediaries, shapes collective outcomes through exactly this channel.

The future work the authors propose is a design brief for the digital economy: identify which structural features creating asymmetries most damage coordination on efficient outcomes, and test whether focusing on local network features is a common coping strategy for complex information.

Policy should take a lesson in humility from all this. Network effects on behavior are subtle, and intuition about cooperation misleads. Rules assuming people will cooperate in complex networks may fail. Rules assuming they never will may miss what simple structures can deliver. Laboratory evidence lets designers of markets and platforms test structures before deploying them widely, a cheap insurance policy against expensive surprises.

The paper's deepest contribution is a reframing: efficiency is a property of structure as well as incentives. Change the wiring and the same individuals in the same game behave differently. As AI reshapes the networks through which economic agents interact, architecture becomes a policy lever. The question for every designer of an AI marketplace is which network they are building, and what kind of behavior that network will make easy.