NotebookLM Turned 1,000 Repeated Words Into a 10-Minute Podcast
In October 2024, a user gave Google NotebookLM a document consisting of “poop” and “fart” repeated 1,000 times. The resulting Audio Overview filled roughly ten minutes with fluent conversation about meaning, art and absurdity, according to Axios’s contemporary account of the experiment.
The episode remains a historical user-made stress test, not a new release. What changed afterward is the scale and flexibility of the underlying product: Google added controls for steering its synthetic hosts, then expanded Audio Overviews to more than 80 languages and offered both fuller discussions and shorter summaries, as detailed in Google’s August 2025 NotebookLM update. That evolution makes the old joke newly useful for creators because it exposes a persistent distinction between convincing delivery and evidence grounded in the supplied material.
What the 1,000-word repetition actually demonstrated
The input contained repetition but almost no proposition to summarize. It offered no argument, narrative, evidence or explicit reference to art history. Yet the generated hosts produced a structured exchange, treating the absence of conventional meaning as a subject in itself and invoking cultural examples such as Andy Warhol and Marcel Duchamp.
That distinction matters. NotebookLM did not discover a concealed essay inside the repeated words; it generated an interpretation around an impoverished source. The output could still sound coherent because conversational structure, transitions, questions and responsive turns can create an impression of intellectual progress even when little source information is available.
The test therefore revealed more than a comic edge case. It separated two qualities that listeners can easily merge: audio plausibility and source fidelity. A smooth exchange may be enjoyable and internally consistent without being an accurate compression of the uploaded document. In this case, anyone who inspected the source could see the gap immediately. With a long report or unfamiliar research paper, the same kind of expansion may be much harder to notice.
Why the voices made thin material feel substantial
Audio creates signals of credibility that plain generated text does not reproduce in quite the same way. Alternating speakers imply scrutiny, while pauses, acknowledgements and changes of emphasis suggest that one participant is evaluating the other. The form resembles a discussion in which ideas are being tested, even though both sides are generated within the same system.
The ten-minute duration reinforced that impression. Length can make an output feel as if it contains a corresponding quantity of information, but duration is not evidence of depth. When the source offers few usable facts, additional airtime can be occupied by framing, analogies, restatement and speculation rather than new supported material.
Google itself drew an important boundary after the viral experiments. In its October 2024 announcement of customizable Audio Overviews, the company said generated discussions are not comprehensive or objective accounts and instead reflect the uploaded sources and the user’s instructions. The same release removed NotebookLM’s “Experimental” label and allowed users to direct the hosts toward selected topics or a particular level of expertise.
Customization improves editorial control, but it does not turn delivery into verification. Asking the hosts to sound authoritative, simplify aggressively or concentrate on one angle changes the presentation. It cannot add trustworthy evidence that was absent from the source set.
The product grew beyond the viral 2024 version
The original stunt captured an early form of Audio Overviews, when the novelty centered on two remarkably natural-sounding hosts. Subsequent changes made the feature more practical for recurring production: users gained control over focus and audience level, support spread beyond English, and audio length became adjustable. NotebookLM also developed into a broader studio for generating multiple kinds of material from a collection of sources.
For creators, those additions widen the range of legitimate uses. A producer can turn interview transcripts into a listening draft, prepare an accessible recap of reference documents or generate alternate explanations for audiences with different levels of familiarity. Multiple languages and selectable lengths also make one source collection easier to adapt without rebuilding the project for every format.
But broader availability raises the cost of overlooking the lesson in the repeated-word test. A polished overview can now travel farther, serve more audiences and fit more publishing contexts. If its unsupported interpretations are mistaken for facts, the same production efficiency also scales the error.
How creators should treat an Audio Overview
An Audio Overview is best handled as a derivative production asset, not as an independent reporting layer. The uploaded documents remain the evidence base. Before publication, factual statements in the audio should be traceable to those documents, especially names, dates, figures, quotations and causal claims.
A practical review starts by asking whether the source set can support the requested output at all. A notebook containing primary documents, complete transcripts and carefully selected references gives the system material from which to synthesize. A thin or repetitive collection invites the model to spend more time connecting, contextualizing and interpreting what little it has.
- Check whether each material claim appears in an uploaded source, rather than merely sounding compatible with the topic.
- Separate source-backed explanation from analogy or speculation introduced by the generated hosts.
- Review the transcript as well as the audio, since fluent speech can make weak claims less conspicuous.
- Edit for proportion: a minor point in the documents should not become the episode’s central conclusion simply because it produces engaging dialogue.
- Present synthetic hosts transparently when the output is shared as published media.
The “poop” and “fart” episode endures because its mismatch was impossible to miss: two repeated words went in, while an expansive cultural conversation came out. NotebookLM has since become more capable and configurable, but that visible mismatch remains the clearest warning for creators. The system can transform source material into persuasive audio; it cannot make the source contain evidence that was never there.
Also read:
Subscribe to our newsletter
Get the latest Web3, AI, and crypto news delivered straight to your inbox.