Subsidiarity

In the AI Story

Applied to AI governance, the principle exposes a distinction the current industry structure ignores: the distinction between functions that genuinely require centralization (computational infrastructure for training frontier models) and functions that do not (governance of the tools built on that infrastructure, terms of service, pricing, capability decisions, data practices). The centralization of the first is a technical necessity that current technology cannot avoid. The centralization of the second is a political choice that current institutional arrangements embed without examining.

The principle does not demand that AI infrastructure shrink. It demands that governance of that infrastructure be organized according to subsidiarity — with centralized functions limited to what genuinely requires centralization, and with decisions affecting users pushed closer to users through democratic mechanisms, cooperative ownership, or regulatory frameworks that prevent the concentration of governance in the same institutions that provide the infrastructure.

The principle has practical force as a diagnostic instrument. Asked of any AI governance arrangement — 'is this decision being made at the lowest level competent to make it?' — it produces clear verdicts that the current industry structure systematically fails. Decisions about which data to train on, which outputs to filter, which capabilities to release, which users to permit, are being made at the highest level (corporate executives, sometimes with regulatory consultation) with almost no participation from the communities affected. Subsidiarity's demand is not to move these decisions to individuals but to move them to the smallest organizational unit that can competently make them — which is often a community, a profession, a regulatory body representing affected populations, rather than a corporate board.

The principle's limitation is that it specifies the goal without specifying the mechanism. Once subsidiarity is accepted as the criterion, the question of how to realize it at scale remains open. Cooperative ownership, regulatory frameworks, public AI infrastructure, data trusts, community governance — each is a candidate mechanism, and each has partial viability but none provides a complete solution. The work of building subsidiarity-compatible AI governance is institutional work that has barely begun.

Origin

The principle was articulated in contemporary form in Pius XI's 1931 encyclical Quadragesimo Anno, which drew on earlier Catholic social teaching. It had intellectual roots in Aristotelian political philosophy, Aquinas's treatment of the common good, and the Reformed tradition's concern with sphere sovereignty.

Schumacher, who converted to Catholicism in 1971, adopted the principle into his economic framework as the structural complement to appropriate scale. It has been further developed in European Union jurisprudence (Maastricht Treaty, 1992), in federalism debates across constitutional democracies, and in contemporary work on platform governance by scholars including Yochai Benkler, Frank Pasquale, and K. Sabeel Rahman.

Key Ideas

Lowest level competent. The decision belongs to the smallest unit that can genuinely make it; moving it higher requires justification that the smaller unit cannot perform the function.

Protects autonomy from absorption. The principle is not anti-authority; it is against illegitimate absorption of functions the smaller unit could perform, which is how large institutions encroach on smaller ones by default.

Applied to AI: the distinction between centralized infrastructure (legitimately so) and centralized governance (not legitimately so) is the diagnostic the principle provides.

Mechanism vs. criterion. The principle specifies what governance should look like; the question of how to realize it at scale remains open and is the proper domain of institutional construction.

Not the same as decentralization. Subsidiarity does not require pushing decisions to the smallest possible unit; it requires pushing them to the smallest unit that can competently make them, which is often collective rather than individual.