Answers are grounded in this paper and cite their sources.
Generative artificial intelligence (GenAI) systems are increasingly designed to initiate, maintain, extend, and participate in conversations with users. Although substantial research has investigated AI refusal mechanisms, existing approaches have predominantly concentrated on declining individual requests, preventing harmful outputs, enforcing content policies, and managing model non-compliance. These approaches do not fully conceptualize conversation itself as an activity that either an AI system or a human participant may voluntarily decline to enter, continue, or maintain. This paper introduces Generative AI Conversation Refusal (GAICR), a novel conceptual framework that establishes refusal of conversational participation as an independent dimension of generative AI interaction. To the best of our knowledge, this paper presents the first unified framework specifically organized around bidirectional and unidirectional conversation refusal, integrating both AI-initiated and user-initiated refusal within a common conceptual architecture. The proposed framework distinguishes four fundamental dimensions: refusal initiator, refusal target, conversational scope, and refusal duration. It further identifies multiple operational forms, including invitation refusal, conversation initiation refusal, topic refusal, turn-level refusal, interaction suspension, conversational termination, participant exclusion, and refusal of particular communication modalities. The framework introduces a structured refusal lifecycle encompassing refusal recognition, intention confirmation, scope determination, refusal communication, interaction-state transition, and potential resumption. Importantly, GAICR distinguishes conversational refusal from conventional safety refusal, technical inability, temporary unavailability, and ordinary conversational disengagement. By recognizing the legitimacy of both human and AI participation boundaries, this research proposes a conceptual foundation for more autonomous, controllable, transparent, and mutually respectful conversational systems. The paper also establishes a research agenda for implementation, evaluation, interaction governance, and future empirical validation.
No takes yet. Share an insight, caveat, or question.
Jincheng Zhang (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: