Gated Uptake in AIUX: Tone, Context, and Speaker-Model Debt in Human-AI Communication

AI systems are increasingly becoming persistent speakers: tutors, assistants, companions, customer-support agents, workplace collaborators, institutional representatives, and autonomous agents communicating across repeated encounters. Yet much AIUX design still treats interaction as response generation: the user says something, the system produces content, and quality is evaluated through relevance, correctness, helpfulness, safety, or style. This paper argues that this model is insufficient. Human receivers do not process utterances as isolated content. They infer intent through tone and context, judge sincerity, and admit or refuse uptake through an accumulated model of the speaker. This paper proposes gated uptake as an AIUX architecture for effective human–AI communication. Context frames interpretation, tone carries relational intent, and content specifies the local semantic payload; but none of these directly produces social effect. The inferred intent must pass through a speaker-model gate shaped by prior interaction, perceived sincerity, refusal history, apology credibility, institutional constraint, category assignment, and available context. AI systems therefore accumulate speaker-model debt: every generated warmth, refusal, hedge, apology, caveat, error, and tone choice modifies the conditions under which future outputs will be received. Once the user’s model of the AI closes, even correct content may no longer land. The paper develops three AIUX consequences of this architecture. First, AI sincerity is structurally ambiguous because generated tone may track persona, policy, brand voice, developer intention, institutional incentive, or nothing stable at all. Second, AI systems often speak in an expression-frame of warmth, care, or helpfulness while institutional material intentions shape what they actually do; the detectable gap between these layers is a major source of performed-feeling AI communication. Third, users often protect prior category models of “AI” by treating surprising systems as reclassified exceptions rather than updating the category itself. The contribution is architectural rather than component-novel: speech-act theory, pragmatics, prosody, trust in automation, HCI, social cognition, and attachment theory each contain pieces of the mechanism. AIUX now requires those pieces to be assembled into a design-level model of uptake.