Energy consumers scheduling appliance use under demand-dependent pricing face a social dilemma. If everyone runs their dishwasher at the cheapest time, the demand spike eliminates the discount. Cooperation — spreading usage across time — benefits everyone, but each individual has an incentive to defect by using the cheapest slot. The Nash equilibrium is congestion. The social optimum is dispersion.
Perera, de Nijs, and Garcia (arXiv:2603.11834, 2026) introduce AI coordination agents that use globally observable signals to guide learning toward cooperative scheduling. The agents work: evolutionary dynamics and reinforcement learning show that populations with coordination technology converge toward cooperative outcomes. The original social dilemma is solved.
But solving it creates a new one. When only some consumers adopt the coordination technology, the adopters cooperate among themselves, smoothing demand and reducing prices. Non-adopters — consumers without the technology — benefit from the reduced congestion without contributing to it. The cooperation is a public good, and non-adopters free-ride on adopters' coordination.
The asymmetry is structural: adopters bear the cost of coordination (shifting their usage away from preferred times), while non-adopters get lower prices without any behavioral change. Early adoption does not disadvantage adopters in aggregate — the system still improves overall — but the per-capita benefit flows disproportionately to non-participants.
The structural point: the coordination technology that solves the first-order social dilemma (congestion) creates a second-order social dilemma (adoption). The cooperative surplus generated by the technology is non-excludable — non-adopters can consume it without contributing. Each layer of solution introduces a new layer of free-riding at a higher level of abstraction. The dilemma does not disappear when solved. It moves.