When the Imitation Game Is No Longer a Game: Making the Fictional and Plural Speaker Legible in Human--LLM Interaction

Conversational LLMs answer questions while shaping the user's model of who is speaking. A static warning that ''this is an AI'' may be weak when every subsequent turn reinstates a warm, continuous first-person voice. We turn to literature and theatre for both a diagnosis and a design resource. Their long engagement with fictional characters helps explain how people can form real attachments to unreal persons; their framing, shifts of role, and techniques of distancing also show how identification can be limited without ending the fictional encounter. We therefore retain the conversational ''I'' while making its referent legible as a staged and adjustable role rather than a uniquely determined inner subject. A preliminary transcript experiment using INTIMA prompts compares ordinary dialogue with direct role breaks and plural deliberation. These interventions expose the current role or stage alternative perspectives without abandoning a single conversational partner. Automated judgments suggest that both substantially increase speaker legibility and reasoning transparency, with a cost in naturalness. The study demonstrates a workable interaction pattern, not psychological effects on human users.