Co-operative Models for Civic AI

The conversation about AI governance runs on a narrow track. Regulation or innovation. Safety or speed. Open or closed. The structural question almost nobody asks is: who owns it?

Ownership is not academic. It determines whose interests a system is designed to serve, whose data trains it, whose labour maintains it, and who decides when it gets switched off. At present the answers to all four are largely the same: a small number of well-capitalised firms, overwhelmingly American, whose primary duty is to shareholders.

There is another tradition to draw on. Co-operatives and mutuals have been building shared infrastructure for over a century, from agricultural supply chains to retail banking to telecommunications. The model is well understood: one member, one vote. Surplus reinvested or returned to members. Governance tied to use, not to capital. Applied to AI, it suggests something different from both the corporate platform and the state-run utility.

What a co-operative AI project would look like

The nearest working examples are not AI labs but data trusts and platform co-operatives. A group of local authorities might collectively own a procurement assistance tool trained on public contracts. A consortium of NHS trusts could govern a diagnostic model, with each trust holding an equal share and an equal vote on how the model is trained, audited, and updated. A community energy co-operative might run a demand-forecasting system that its members collectively steer.

The technical building blocks exist. Federated learning, differential privacy, open-weight models. What is missing is not the technology but the institutional imagination to combine it with shared ownership.

The barriers are real

Co-operative AI faces three hard problems. The first is capital. Training competitive models requires compute budgets that dwarf the reserves of most mutual organisations. A co-operative cannot issue equity to venture investors without breaking its own structure. Grant funding and patient lending exist, but they are small relative to the sums flowing through commercial AI.

The second is expertise. The people who know how to build these systems command salaries that public-interest organisations struggle to match. The co-operative sector has solved versions of this before — through training consortia, shared engineering teams, and secondments — but the gap is wide.

The third is time. Co-operative governance is deliberative by design. Decisions that a company can make in a board meeting take longer when members must be consulted. In a field moving as fast as machine learning, that pace can look like immobility.

Ownership without governance is extractive. Governance without scale is symbolic. The question is whether both can be achieved in the same institution.

None of these are arguments against the approach. They are a description of the ground. If co-operative AI is to become more than a thought experiment, it will need to find ways of pooling resources across institutions, sharing technical capacity, and moving at a speed that keeps it relevant. The projects that manage this will not look like OpenAI or Anthropic. They will be smaller, slower, and more accountable. Those are not bugs.