How groups govern a shared resource
Shared AI resources, from a co-op's computing to a city's data, can be governed by the people who use them rather than by one owner or one coordinator. Communities have done this with pastures, water and fisheries for centuries, and the rules that made those commons last apply to AI too.
What the commons research found
In 1968 Garrett Hardin argued that people sharing a pasture will overgraze it, so a shared resource must be either privatised or run by a state.1 Elinor Ostrom spent her career documenting communities that did neither and kept their resource healthy for generations. In the alpine village of Törbel, for example, shared pastures have been managed under village rules since at least the thirteenth century.2 She found that the long-lived ones shared a set of design features:
- clear boundaries on who is a member and what the resource is;
- rules that fit local conditions, which the members themselves can change;
- monitoring by people accountable to the members;
- sanctions that start small and escalate;
- cheap, local ways to settle disputes;
- outside authorities that respect the group's right to organise;
- for large systems, many nested groups rather than one centre.3
Why this matters for AI
Two things are new. People increasingly share AI resources, such as a co-op's models and computing, a city's data, or a community's moderation rules. And AI agents are starting to share resources among themselves. In a simulated fishery, most language-model agents over-harvested the common pool; only the most capable models kept it alive, and the best survival rate was below 54%.4 When one agent was given more power than the others, cooperation broke down further, with survival rates falling by up to 87%.5
What we propose
- Rules the members can amend, in small steps. Keep the rulebook versioned, so a proposed change is a visible edit to one rule, tried for a fixed period, and kept or reverted. Communities avoid both frozen rules and late-night rewrites where the loudest voice wins.
- Authority that rotates and expires. Whoever coordinates, person or agent, holds the role for a limited term and has to be renewed (see AI agents should hold permissions that expire). A simulation testbed now exists where human and machine participants choose and change their own governance arrangements under enforced rules, so these designs can be tried before they are used.6
- Many small, overlapping groups instead of one coordinator. A single coordinator sees only what reaches its desk and accumulates power nobody can check. Game-theoretic work shows that members' ability to leave keeps a delegate honest only up to a certain group size; beyond it, the delegate can exploit members without losing them.7
- A cheap exit. Members can take their data and leave, or fork the rules and start a new group. The threat of exit disciplines whoever is in charge.8
- Monitoring that can see agents coordinating. Agents can collude in ways transcripts don't show; researchers are testing detectors that read the models' internal activations.9
The objection
Central coordination is sometimes simply more efficient. In one synthetic economy benchmark, a central planner consistently beat market and hybrid mechanisms, though the author cautions that the design favoured the planner and that the result does not establish planning's superiority.10 A centre can help. It still needs term limits, monitoring and an exit, and most of the time many smaller groups are safer.
What's unsolved
- Verifying that each member is one real person, without a surveillance register.
- Detecting agents that coordinate in ways neither their messages nor their actions reveal.
- How large a group can grow before exit stops keeping its delegates accountable, for groups that include AI agents.
Related: Hearing how much people care, not just how loud they shout, on how a group can decide what it wants.
Sources
1. Garrett Hardin, "The Tragedy of the Commons," Science 162, no. 3859 (1968): 1243–1248, doi:10.1126/science.162.3859.1243.
2. Elinor Ostrom, Governing the Commons: The Evolution of Institutions for Collective Action (Cambridge University Press, 1990), doi:10.1017/CBO9780511807763.
3. Elinor Ostrom, "Beyond Markets and States: Polycentric Governance of Complex Economic Systems," American Economic Review 100, no. 3 (2010): 641–672, doi:10.1257/aer.100.3.641.
4. Giorgio Piatti, Zhijing Jin, Max Kleiman-Weiner, Bernhard Schölkopf et al., "Cooperate or Collapse: Emergence of Sustainable Cooperation in a Society of LLM Agents" (2024), arXiv:2404.16698.
5. Abhilekh Borah, "Bosses, Kings, and the Commons: Cooperation Under Power Asymmetry in LLM Societies" (2026), arXiv:2605.29062.
6. Callum Berger, Jeremy Pitt and Stefan Sarkadi, "A Simulation Testbed for Self-Governing Human-Machine Institutions" (2026), arXiv:2610.36989.
7. Arin Mukhopadhyay, Aris Filos-Ratsikas and Panayiotis Danassis, "The Delegation Dilemma: Economies of Scale, Accountability, and Polycentric Governance" (2026), arXiv:2610.21746.
8. Vincent Ostrom, Charles M. Tiebout and Robert Warren, "The Organization of Government in Metropolitan Areas: A Theoretical Inquiry," American Political Science Review 55, no. 4 (1961): 831–842, doi:10.2307/1952530.
9. Aaron Rose, Carissa Cullen, Sahar Abdelnabi, Philip Torr et al., "Detecting Multi-Agent Collusion Through Multi-Agent Interpretability" (2026), arXiv:2604.01151.
10. Ricardo Alonzo Fernández Salguero, "Testing Centralized and Polycentric Computational Planning" (2026), arXiv:2606.19214.
