Personalization Fosters Pluralistic Alignment under Competition

The researchers prove in a Stackelberg game model that when multiple self-interested AI providers compete, user-level personalization can, under the Weak Market Alignment condition, let each user obtain results comparable to a fully aligned public model, i.e., personalization can foster pluralistic alignment even when providers are self-interested. In the model, providers act as leaders committing to conversation policies, and users act as followers choosing which model to use, how to converse, and how to act. In contrast, if providers can only deploy a single anonymous policy, under the same conditions there exists an equilibrium in which behavior provides no information; the paper additionally gives a stronger alignment condition that guarantees each user achieves optimal utility in the anonymous case.