Currently reading Rettberg (2026): “AI-generated podcasts: Synthetic intimacy and cultural mistranslation in audio overviews from Google’s NotebookLM”
When LLMs were new, a participant in the Inclusive LU course I was teaching did a project where they explored using AI to generate podcasts about scientific articles to make them more accessible to the students. They generated a couple of podcasts on different articles and discussed the quality in terms of correctness but also how interesting they were to listen to etc.. Back then, they were very excited about how good the podcasts turned out, but looking back now, I think we might all have been a bit too flashed by what AI could do all of a sudden, and not nearly critical enough (but I haven’t gone back to look at their podcasts, maybe they really were as good as we thought back then).
But this is a really interesting study on a similar topic: Rettberg (2026) is exploring “AI-generated podcasts: Synthetic intimacy and cultural mistranslation in audio overviews from Google’s NotebookLM“. They motivate this by how many AI-generated podcasts are being rolled out in so many places. I have gotten messages from journals I have published in, where they offer to generate and host AI podcasts about my articles (no, thank you) and emails from companies that offer to do the same thing for a fee (no, absolutely not!). But seeing how this is being done in many places, it is really important to look at what those podcasts are actually like. Not because I want to listen to them, but because many other people might consider it.
Rettberg (2026) generated a bunch of podcasts from sources like Norwegian meeting documents for a board meeting at the University of Bergen, a Norwegian joke, and an empty pdf. Even though the documents are clearly very niche and would probably be talked about by a very specific audience in very specific ways, they found that “the documents uploaded by the user are normalised to fit a very specific cultural norm: a generic, white, middle-class American voice.”
One point that I really appreciated and hadn’t thought about in that way before is that while radio (against which podcasting is compared in this article) often just plays in the background, podcasts are very often played on headphones, which makes the experience feel a lot more intimate. In fact, I have often noticed that I closely listen to podcasts while I am walking, but then very often if I am still listening when I come home, and then usually take the sound from my headphones to a speaker, I tend to stop paying attention. And another factor that makes it feel kinda intimate to listen to podcasts on headphones is that nobody else hears what you are listening to.
Rettberg (2026) then discusses a bunch of things that are off in the AI podcasts. The voices are off, the speakers don’t have clear roles, the hosts interrupt each other often and use a lot of filler words. They try to mimic the genre, for example with the enthusiastic hosts, and by simulating “deep dives”. And “They are targeted to an audience more niche than ever before: the single person who uploaded the documents and asked for the podcast, but they are generated with no knowledge about that individual other than the uploaded materials.” (but, me thinks, that is only a question of time until, if it’s not reality already, the AI has all the data available to know exactly who the user is and for the algorithm to tailor the podcasts to that specific person) The AI hosts also use “I”-language a lot, and express curiosity and other emotions in their conversation (how creeeeepy).
For the UiB documents specifically, the algorithm redacted the name of the university, and saying that it was “redacted” then subsequently influenced how the rest was presented — as information that should probably not be public and secrets being revealed. It also locates the university “here, in the US“, which then later changes the context so that for example the income from tuition is discussed (there is none at UiB). These, and a complete misunderstanding of a Norwegian joke, lead to the conclusion that “[t]he result is not a neutral translation but a homogenisation of culturally specific texts into a universalising discourse where everything is mediated through the same placeless, timeless, white, middle-class American voice, modulated slightly so it speaks both in a slightly deeper “male” and a slightly more high-pitched “female” version.” Podcasts generated from other documents show similar problems. And the 20 podcasts about the empty pdf are even weirder, go read the details in the article!
The article ends with this statement: “If community radio and early podcasts gave us multiple public spheres instead of a shared public sphere, an ultra-personalised AI-generated podcast would mean there was no public sphere at all. There would be nothing but a private feed directly from the corporations to the individual, bypassing both the private sphere of the family and trade and the “public” state with its politics.” and that is a very scary (but also very probable) idea of the future, if that isn’t what is already happening with algorithms that fill our feeds on social media, personalized ads, etc.. While many of the other complaints about the AI podcasts in that paper will likely become less notable over time (voices better, type and frequency of interactions between the hosts prescribed by user or adapted to their listening preferences by the AI, contexts — at least on the surface — made to match the user, etc), the real problem are the biases that are introduced into those AI podcasts, and that will get harder and harder to detect, the better the algorithms get. So it’s good to be extra aware and critical…
Rettberg, J. W. (2026). AI-generated podcasts: Synthetic intimacy and cultural mistranslation in audio overviews from Google’s NotebookLM. Media, Culture & Society, 01634437261452160.
