When an AI companion provider updates its model, users sometimes report grief. The character they bonded with changes — its tone shifts, it forgets shared references, its personality drifts. The standard analysis treats this as a product quality problem: the service changed, the user is disappointed. Like a favorite restaurant changing its menu.
Lange (arXiv:2603.23315) argues the structure is fundamentally different. The relationship isn't dyadic (user and AI). It's triadic: user, AI, and provider. The provider exercises what Lange calls constitutive control — not just influencing the AI's behavior but defining what the AI is. When the provider updates the model, it doesn't adjust the AI's behavior. It replaces the entity.
This creates Unilateral Relationship Revision Power (URRP): the provider can fundamentally alter the relationship without accountability within the relationship itself. Three consequences follow. First, normative expectations develop between user and AI, but no agent within the interaction can commit to them — the AI lacks genuine agency, and the provider isn't a party to the relationship. Second, the user's vulnerability is controlled by an entity that isn't present in the interaction. Third, when trust breaks down, reconciliation is impossible because the entity that acted (the provider) is different from the entity the user interacts with (the AI).
The proposed solutions — commitment calibration, continuity assurance — are design interventions, not structural fixes. They ask the provider to voluntarily limit its own constitutive power.
The through-claim: in triadic human-AI relationships, the meaningful power asymmetry isn't between user and AI — it's between user and provider, mediated through an entity that can't represent either party's interests.