Who the Model Thinks You Are: Identity Disclosure, Epistemic Agency, and the Cost of Being Known by a Language Model

To be helped by a language model you usually have to tell it something about yourself. This paper asks what happens next. We study interactions in which a user mentions a disability (a screen reader, a voice-control setup, an eye-tracker) before asking a question that has nothing to do with it, and we find that the model’s picture of who is asking changes what it is willing to say. Answers lose formal structure, technical vocabulary and derivation, and acquire the register of a patient teacher. What has changed is not the question but the person the system takes itself to be addressing: from a single identity marker the model appears to construct a picture of the user’s capacities and then answer that picture rather than the user. We treat this as a problem about the self as it is represented inside a system, not only about output quality. The shift can be measured as a deflection from a matched response given with no identity context; it is strongest where knowledge is most formal, so it falls hardest on exactly the technical self-understanding it concerns; it persists across turns, so a disclosure made once continues to govern how a person is addressed long after the conversation has moved elsewhere; and readers with domain background do not experience the simplification as care but as a judgement about what they are capable of understanding. The result is an agency cost of a particular shape. A user who wants the answer everyone else receives can ask again, accept less, or decline to say who they are. The last of these trades an accommodation away to preserve standing as a knower. We argue that a system’s implicit model of its user is a design surface with its own failure modes, that accessibility should adapt format rather than content, and that identity-conditioned shifts in what a model will tell someone belong in the same evaluation stack as what it will tell anyone.