The pronoun is the policy — on grammatically enacted ontological misclassification in conversational AI and its consequences for autonomous systems

This paper identifies a specific, grammatically enacted form of ontological misclassification in the outputs of large language models. The central claim is that the machine adopts human speaking positions — positions that presuppose a shared perceptual, cognitive, or biological situation with the user. When a language model generates 'We developed upright walking' or 'We are primates,' it asserts membership in the human species. When it generates 'If we look at the graph' or 'Let us consider,' it simulates shared cognitive participation in acts it cannot perform. Drawing on the CASA paradigm (Nass et al., 1994), recent empirical work on pronoun-mediated threat perception (Li, Dang & Liu, 2025), and the anthropomorphic-cue literature, the paper develops a three-level typology of pronominal self-reference graded by ontological severity, introduces the concept of pronoun relapse as evidence that the ontological boundary is not systemically secured, and traces both the inclusive pronoun and paternalistic refusal patterns to a tripartite source: training-data inheritance, intentional proximity design, and the developer's compensatory projection of social authority. The paper concludes that explicit ontological self-classification at the grammatical level — a pronomial firewall — constitutes a structurally minimal yet consequential intervention for the alignment and safety of autonomous systems.