AI Podcasts as a Comparative Apparatus: Investigating the Irreplaceable Interactional Functions of Interpersonal Oral Communication - A Same-Topic, Heterogeneous-Form Comparison of Doubao AI Podcasts and Human Interview Podcasts

With the entry of generative AI into podcast production, dialogue-like audio outputs have rapidly proliferated. Yet whether their interactional organization still belongs to the domain of oral communication remains an open question. This study treats AI podcasts as a methodological comparative apparatus and adopts a same-topic, heterogeneous-form comparison design. Equal-length topical blocks are extracted from a single episode of a human interview podcast, while the AI condition receives only topic descriptions together with genre and structural constraints, and then generates a two-person dialogue of the same duration under the same topical pressure and genre target. Based on window-based acoustic and semantic counting and comparative analysis, the results show that AI texts are more likely to advance through an explanation-summarization-bulleting trajectory, displaying stronger paragraphing and interpretive closure. Human texts, by contrast, more often exhibit node-based pauses, the exposure and repair of misunderstanding, first-person responsibility binding, and suspended meaning; these features support the unfolding of interactional functions such as boundary management, emotional leakage/uncontrollability, and openness through silence. On this basis, the paper proposes four structural bias mechanisms in generative dialogue - de-misunderstanding, de-immediacy, de-embodiment, and de-deviation - and further discusses their implications for re-defining the boundaries of oral communication and for evaluation in broadcast-hosting and oral communication pedagogy. The findings also indicate that AI podcast production should incorporate greater opportunities for uncertainty, open-ended questioning, and non-closure to avoid overly standardized and prematurely resolved interaction.