Experimentation in Federal Systems

GAICIL Insights

Federal systems are often praised as laboratories of democracy, one of the most flattering ideas in political science.¹ Experimentation in Federal Systems, by Steven Callander and Bård Harstad, tests it formally and finds the laboratories may be sabotaging each other. In a model where heterogeneous districts choose both whether to experiment and which policies to try, informational spillovers mean that in the first best districts should converge in their policy choices. Under decentralized authority, the equilibrium predicts the opposite.

The culprit is free-riding. Each district can learn from others' experiments, so each prefers to let others bear the costs. To discourage free-riding, a district may choose a policy less useful to others, particularly when districts are similar. Policies end up more different than districts' ideal points would suggest, inefficiently minimizing the informational spillovers that make federal experimentation valuable. Heterogeneity turns out beneficial because it discourages free-riding. Districts make their experiments deliberately hard to copy.

The remedy is elegant: dynamic federalism, where the central government harmonizes policy only after districts have experimented. This progressive concentration of power induces a policy tournament, raising the incentive to experiment and encouraging convergence. Even a very clumsy central government can play this role. With harmonization at the end, each district may face a stronger incentive to experiment, and positions may converge as in the first best. A center with modest skill can still do the job by arriving late and picking a winner.

The workhorse model is simple and tractable and has been extended to multiple districts and periods, coordination benefits, transparency, imperfect correlation, transfers, prizes and intellectual property rights. Many extensions remain, and the authors note that the prominence of policy experimentation in discourse has not been matched by formal understanding.

AI regulation is a federal laboratory in real time. National governments, states and provinces, and international bodies are writing AI rules simultaneously. The paper suggests this laboratory may not work as hoped. Jurisdictions may choose divergent approaches partly to avoid others free-riding on their experience, reducing the information each gains from the rest. The patchwork of AI laws may be partly strategic.

The remedy translates directly. Jurisdictions first experiment with different approaches, followed by later harmonization at a higher level. The prospect of eventual harmonization turns experimentation into a tournament, where jurisdictions compete to have their approach adopted more widely. This dynamic could beat both immediate centralization and permanent fragmentation. A promise to harmonize in a few years may be a valuable policy design.

Heterogeneity is an asset here. Jurisdictions face different conditions and priorities. Differences encourage genuine experimentation and reduce free-riding. AI policy may benefit from diverse approaches, provided mechanisms exist to learn from results and converge on what works.

Prizes and intellectual property extensions suggest further tools. Rewarding jurisdictions whose experiments yield valuable information, or letting them capture some benefits of their innovations, could strengthen incentives to experiment. Such mechanisms could support regulatory innovation in AI, where uncertainty about rules' effects is high. Imagine regulatory sandboxes competing for the right to become the template.

Timing is the master variable. Harmonizing too early forecloses experimentation; harmonizing never sacrifices coordination benefits. The optimal design concentrates power progressively, after experimentation has generated information. For fast-moving technologies, the timing of harmonization is a crucial design choice.

Seen whole, the paper's message is plain: experimentation in decentralized systems requires careful institutional design. Left alone, jurisdictions may experiment inefficiently, shrinking the information others can use. A central authority harmonizing after experimentation can restore incentives and promote convergence. For AI, where many governments write rules simultaneously, this framework offers a rigorous basis for governance that learns as it goes.

Those pushing an immediate global AI treaty might start here: and for those defending a permanent patchwork: both may be wrong on timing. Announce now that harmonization comes later, and let the jurisdictions compete to write it.