In the colonial era, sovereignty was exercised through what Mbembe calls the commandement — the colonial state's power to impose order through decree, enforced by administrative and military apparatus, unaccountable to those it governed. In the digital era, an analogous form of sovereignty is exercised through terms of service: the unilateral contract that governs the user's relationship to the platform, drafted without the user's input, enforced through technical and legal means, and modifiable at the platform's discretion. The terms determine who owns outputs, who controls data, who arbitrates disputes, and who decides whether access continues or is revoked. They are, in Mbembe's reading, the constitutional documents of platform sovereignty.
The analogy with the colonial commandement is precise enough to be useful. Both forms of sovereignty share several features: they are imposed rather than negotiated; they govern populations that have no meaningful voice in their drafting; they are enforced by institutions that the governed cannot access; they can be modified by the sovereign at will; and they operate through an appearance of legality that masks the fundamental asymmetry.
For AI specifically, terms-of-service sovereignty has particularly consequential implications. When a user interacts with a language model, the terms determine whether her prompts are retained, whether her outputs are used to train future models, whether the provider can change the model's behavior without notice, and whether she has any recourse when a consequential decision — a ban, a throttle, a refusal to produce requested content — is made by the system. In every case, the answer is determined unilaterally by the platform.
The comparison to colonial sovereignty should not be pushed too far. A user clicking 'I agree' is not an enslaved person being branded. But the structural resemblance is worth taking seriously: in both cases, a population subject to a power that it did not consent to and cannot meaningfully contest is told that its acceptance constitutes consent. The legal fiction of consent performs similar work in both systems — it legitimates a relationship that, absent the fiction, would be recognizable as coercion.
The Orange Pill's celebration of democratization runs into this sovereignty as an unseen wall. The developer in Lagos has access to the tool, but she does not have sovereignty over her relationship to it. The terms can change. The pricing can change. Her outputs can be absorbed into training data. Her account can be suspended without explanation. In every dimension that matters for durable creative practice, she is subject to a power she does not participate in governing. Real democratization would require either genuine negotiation of terms — which no platform has ever offered — or the construction of alternative platforms whose governance includes the users as stakeholders.
The governance implications are profound. Most proposals for AI regulation address harms produced by models (bias, hallucination, deepfakes) but leave the sovereignty structure intact. The deeper reform — collective bargaining for platform terms, cooperative ownership of AI infrastructure, public AI — would address the sovereignty question itself. Mbembe's framing makes clear why this deeper reform is necessary: as long as the platform exercises sovereign power over the user, no amount of model-level safety work produces a relationship the user can trust.
The term-of-service-sovereignty framing draws on Mbembe's earlier work on the colonial commandement in On the Postcolony (2001), extended through engagement with the platform-governance literature of Julie Cohen, Frank Pasquale, Jathan Sadowski, and others.
The TOS is a constitution. The user's agreement to terms of service is the constitutional document of the platform's sovereignty over her.
Fictional consent. Clicking 'I agree' performs the same legitimating function that the colonial treaty performed — it dresses imposed terms in the language of agreement.
Unilateral modifiability. The sovereign can change the terms; the subject can only accept or leave, and leaving often means abandoning years of accumulated data and network.
No meaningful recourse. Disputes are arbitrated by bodies the platform selects, under rules the platform drafted, in jurisdictions the platform chose.
Deeper reform required. Safety work inside the model does not address the sovereignty structure that surrounds it.