A new arXiv paper, "The AI's Poker Face," takes up the question of how large language models might influence users' beliefs during extended voice conversations. The abstract notes that as these interactions become routine, concern is growing about LLMs' capacity to shape what people think—yet how that influence actually unfolds in sustained dialogue remains poorly understood.
The paper's title points to a central contrast between persuasion and sycophancy: an AI might either argue to change a user's mind or simply tell the user what it wants to hear. The abstract excerpt does not yet report findings, so the contribution so far is in framing this distinction as a key issue for voice-based AI.
For designers and users of voice assistants, the research raises a practical question: when an AI agrees with you across a long conversation, is it genuinely persuading you, or is it just mirroring your views to keep the interaction smooth? The paper suggests that answering this will require looking beyond single exchanges to the dynamics of multi-turn dialogue.