Much of CUI research still treats the system as an interlocutor: a conversational other that should appear trustworthy, helpful, and socially adept. This paper argues that the dominance of the interlocutor metaphor has become a constraint because it strictly preserves a self-other boundary. Conversational AI is increasingly used not only for dialogue, but to externalize tentative thoughts, offload cognitive work, and support ongoing reasoning. In practice, these systems often function less as independent external interlocutors than as components of the user’s cognitive process. Large language models can imitate the surface patterns of dialogue, but they do not share the embodied and situated conditions through which speakers establish meaning. Designing CUIs primarily as artificial partners therefore risks optimizing the social illusion rather than addressing the deeper problem of grounding. Drawing on extended cognition theory, this provocation suggests that CUIs should be understood not as quasi-persons but as cognitive extensions of their users. Rather than simulating an independent conversational mind, such systems would couple directly to human cognition and support thinking from within. We propose that passive brain-computer interfaces offer one possible route: they could allow CUIs to adapt to signals like cognitive effort or agreement. The aim is not greater human-likeness, but tighter coupling with the human mind.