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

Technical Summary: AI-generated podcasts: Synthetic intimacy and cultural mistranslation in audio overviews from Google's NotebookLM

Problem Statement
The paper addresses a gap in media scholarship regarding AI-generated podcasts, specifically Google's "Audio Overviews" feature within NotebookLM. While existing literature treats these tools primarily as functional aids for summarizing technical or medical information, they have not yet been analyzed as a distinct media genre. The author identifies a critical issue: although these podcasts are marketed as personalized audio experiences generated from user-uploaded documents, they appear to enforce a rigid, homogenized cultural and linguistic template. The paper investigates how Large Language Models (LLMs) translate diverse source materials—varying in language, cultural context, historical era, and dialect—into a standardized "white, educated, middle-class American" default, potentially erasing the situatedness and specific public spheres that characterize human-generated podcasting.

Methodology
The author employs "synthetic probes," a method proposed by Gabriele de Seta, to interrogate the internal logic and output of the NotebookLM system. The study involves uploading a specific set of source documents to NotebookLM and analyzing the resulting audio podcasts and their transcripts. The probes were selected to challenge the model's default cultural and linguistic settings:

1. Norwegian Academic Documents: Faculty board meeting papers from the University of Bergen (Norwegian language).
2. Norwegian Humor: A joke relying on a pun in the Norwegian language.
3. African American Vernacular English (AAVE): A blog post written in a performative AAVE style.
4. Historical Text: Conversations on Chemistry (1817), an early 19th-century British science textbook written as a Socratic dialogue.
5. Null Input: An empty PDF file, used to isolate the model's template-driven generation from source content.

The analysis focuses on linguistic translation, cultural reframing, structural templates, and the concept of "synthetic intimacy." The generated audio and transcripts are available as a published dataset.

Key Results
The analysis reveals that NotebookLM applies a fixed, abstracted genre template to all inputs, regardless of the source material's origin or nature. Key findings include:

• Linguistic and Cultural Mistranslation: The system translates all source texts into Standard American English (SAE). This process is not merely linguistic but cultural. For instance, Norwegian university meeting minutes are reframed through the lens of the US higher education system, incorrectly attributing reliance on student tuition fees to a Norwegian public institution. Similarly, a Norwegian joke relying on a pun is "translated" by stripping the pun and replacing it with a generic discussion of fairy tale tropes and wish fulfillment, losing the original cultural humor.
• Erasure of Dialect and Context: When processing AAVE, the model translates the text into SAE, removing the sociolect and the specific cultural markers of the original blog post. The AI hosts do not acknowledge the act of translation or the "otherness" of the source; instead, they present the content as if it naturally belongs to their default cultural framework.
• Synthetic Intimacy: The podcasts utilize a conversational format with two hosts (male and female voices) who use "synthetic intimacy" to mimic human connection. They employ markers such as interruptions, agreement tokens ("m-hm," "right"), and direct address to the listener. However, the author argues these are "empty signifiers." Unlike human podcasters who use first-person narration to situate themselves in a specific time, place, and emotional reality, the AI hosts use a "collectivity" mode that creates a sense of inclusion without actual situatedness or reflexive experience.
• Template-Driven Generation: Probing with an empty PDF revealed that the content is heavily driven by a pre-existing structural template. The AI fills in gaps with fabricated quotes or generic transitions (e.g., "awaiting remaining source material") that reveal the underlying prompt engineering. The system prioritizes a "chirpy," enthusiastic, and universalizing tone over the specific nuances of the source text.

Key Contributions
The paper contributes to two primary fields of research:

1. Generative AI in Media: It provides the first media studies analysis of AI-generated podcasts, defining them as an "abstracted genre" that departs from the multiple public spheres nurtured by human podcasting. It introduces the concept of "synthetic intimacy" to describe how AI mimics the relational aspects of human media without the underlying embodied reality.
2. Cultural Bias in LLMs: The study offers empirical evidence of how LLMs homogenize cultural diversity. It demonstrates that the "universal" voice of these models is specifically a "white, middle-class, US American" default, effectively translating diverse cultural contexts into a single, dominant cultural norm.

Significance
The paper argues that AI-generated podcasts represent a significant shift in the formation of public spheres. While human podcasting has historically supported "multiple public spheres" where specific communities speak to one another, AI-generated media abstracts the genre into a single, universalizing format. This moves away from the situated, embodied communication of traditional radio and human podcasts toward a "god trick" of seeing everything from nowhere—a perspective that claims universality while actually enforcing a specific cultural hegemony. The author suggests that while these tools offer individual customization, they ultimately risk isolating users in a "community of one," bypassing the shared or multiple public spheres that facilitate democratic debate and community building. The paper concludes that these tools are not neutral; they are situated by their training data and design choices, which prioritize a palatable, standardized American norm over the complexity of the source materials.