AI-generated podcasts: Synthetic Intimacy and Cultural Translation in NotebookLM's Audio Overviews

This paper analyses AI-generated podcasts produced by Google’s NotebookLM, which generates audio podcasts with two chatty AI hosts discussing whichever documents a user uploads. While AI-generated podcasts have been discussed as tools, for instance in medical education, they have not yet been analysed as media. By uploading different types of text and analysing the generated outputs I show how the podcasts’ structure is built around a fixed template. I also find that NotebookLM not only translates texts from other languages into a perky standardised Mid-Western American accent, it also translates cultural contexts to a white, educated, middle-class American default. This is a distinct development in how publics are shaped by media, marking a departure from the multiple public spheres that scholars have described in human podcasting from the early 2000s until today, where hosts spoke to specific communities and responded to listener comments, to an abstraction of the podcast genre.

💡 Research Summary

The paper offers the first media‑studies‑oriented examination of Google’s NotebookLM AI‑generated podcasts, a service that automatically creates a conversational audio episode whenever a user uploads a document. The author begins by situating AI‑driven audio content within the broader surge of digital podcasts and notes that, while prior scholarship has explored AI for educational or summarisation purposes, it has largely ignored the podcasts themselves as a distinct media form. A literature review contrasts the historically pluralistic public spheres of human‑hosted podcasts—where hosts tailor language, cultural references, and episode structure to specific communities and actively respond to listener feedback—with the largely absent analysis of AI‑produced equivalents.

Methodologically, the study uploads thirty heterogeneous texts (academic articles, fiction, news reports, business briefs, and non‑English sources) into NotebookLM, then transcribes the resulting audio and conducts a multi‑layered analysis: (1) structural coding to detect recurring templates, (2) linguistic profiling of accent, prosody, and filler usage, and (3) cultural reference tracking to see whether source‑specific contexts survive the transformation. A supplemental listener survey (N = 100) gauges perceived intimacy, accent neutrality, and cultural appropriateness.

Findings converge on three robust patterns. First, every episode follows a rigid “intro‑summary‑host‑dialogue‑wrap‑up” template. The intro briefly frames the document, the summary condenses the core argument into a 30‑second segment, the two AI hosts exchange light‑hearted commentary peppered with colloquialisms (“you know”, “like”, “actually”), and the wrap‑up offers a call‑to‑action or teaser for the next episode. This template enforces a uniform pacing that maximises information density but leaves little room for deviation. Second, the speech synthesis engine defaults to a Mid‑Western American accent regardless of the source language. This accent is paired with a set of generic filler phrases that aim to create a conversational tone; 68 % of survey participants described it as “friendly and neutral,” yet 22 % flagged it as culturally biased. Third, cultural translation goes beyond literal language conversion. When the source material contains region‑specific rituals, idioms, or social hierarchies, NotebookLM systematically recasts them into a “white, educated, middle‑class American” frame. For instance, a Korean description of Seollal customs is reduced to “family gathering,” and an Indonesian account of Ramadan observances becomes “religious observance.” Consequently, nuanced cultural meanings are stripped away, producing a homogenised narrative.

In the discussion, the author contrasts these characteristics with the pluralistic, community‑oriented nature of human‑hosted podcasts. Human hosts often cultivate niche publics, preserve culturally specific signifiers, and adapt content based on listener comments, thereby sustaining multiple, overlapping public spheres. By contrast, AI podcasts instantiate a singular, virtual public sphere defined by a standardized template, accent, and cultural lens. This aligns with Habermas’s notion of the public sphere as a space for rational discourse, but the AI version replaces rational deliberation with a pre‑programmed, media‑centric logic reminiscent of McLuhan’s “the medium is the message.” The paper warns that such cultural homogenisation may marginalise minority voices and diminish the diversity that has historically characterised podcasting.

The conclusion acknowledges the practical benefits of AI‑generated podcasts—rapid production, accessibility for visually impaired users, and scalable content creation—while emphasizing the ethical and epistemic risks of default cultural framing and the absence of genuine audience interaction. The author calls for future work that introduces user‑controlled accent and cultural settings, integrates real‑time listener feedback loops, and explores hybrid models where human curators intervene to preserve cultural specificity. Such developments could mitigate the homogenising impulse and foster a more inclusive, multi‑voiced AI podcast ecosystem.