[Author’s Note: Predictions for 2025 will be published on Thursday. This essay came to mind after using NotebookLM last week.]
Last Friday I decided to run an experiment with last Thursday's essay (“Was I Right?...”): I uploaded it to NotebookLM to create a podcast. Within minutes, a 19-minute “podcast” of my essay emerged with an anonymous, AI-generated male-female duo enthusiastically discussing key points from my essay (and you can listen to it on my old PARQOR Substack).
A reader texted me that it sounded just like one or more of the Penske Media-owned podcasts, and better than the ones he had listened to. Those podcasts have human hosts, production budgets and a loose script.
This was an AI-generated, “tightly”-produced “podcast based on an analytical essay, structured according to an ordered sequence of subheadings, and scripted as a debate around what I wrote. Most importantly, it was free.
This podcast production challenges both my essay format and any potential podcast I could create. The irony is not lost on me as someone who writes about all the new and exciting ways AI is challenging the movie and TV formats. I have written about the future of text-to-prompt video but this is essay-to-podcast audio. It is a different "animal".
The whole experience reminded me of something an AI filmmaker said in passing during our conversation the previous week: “a script is a 100-page analog prompt that we give to 300 people and say go and make this film.” NotebookLM took my essay as an "analog prompt" for producing a podcast, even if the essay was not intended to be a script.
That reads and feels different than anything I have discussed with entrepreneurs in the Generative AI space, to date.
Key Takeaway
NotebookLM suggests AI will be less an original storyteller and more a producer of faster, cheaper “performances”. But, a storytelling "performance" is still storytelling, and therein lies the rub.
Total words: 1,000
Total time reading: 4 minutes
The Analog Prompt
It reminds me of something Producer and Actor Ben Affleck said last month:
“AI can write you excellent imitative verse that sounds Elizabethan. It cannot write you Shakespeare. The function of having two actors, or three or four actors, in a room, and the taste to discern and construct that, is something that currently entirely eludes AI's capability, and I think will for a meaningful period of time.”
In other words, AI is bad at storytelling. The AI filmmaker told me this is because AI "lacks that storytelling compression layer that we've got.”
Samir Chaudry of YouTube creator duo Colin and Samir told me something similar in an interview earlier this year: “it is in our DNA to tell stories and to hear stories told” by human storytellers.
NotebookLM seems to reveal a blind spot in these arguments. The podcast generated by NotebookLM may be a better “storyteller” about my analysis than the essay in which I delivered it, or a podcast I could record by myself.
This podcast seemed to have “worked” because my essay had a basic structure to it:
Subheader for each prediction
Emoji symbols reflecting my evaluation of how I did
The discussion
The “script” of the podcast follows that outline in order. It used my website URL to identify where I worked (and it missed the nuance that I do not work at The Information). There are no prompts in the script that tells the AI to make the voices enthusiastic or doubting in different areas.
Nevertheless, it was able to generate a compelling audio “story” out of my essay that sounded like any generic AI-generated male-female podcast hosting duo. In other words, it was AI-generated storytelling with human voices about a story I had told in essay format.
An Important Nuance
There is an important nuance here. Chaudry’s concern about flooding was that AI could generate its own compelling video stories from consumer text prompts. That is certainly still possible.
Here NotebookLM is generating compelling audio from an essay that has been outlined, written, edited and published. Both the input and the output have quality and utility to my readers and target audiences.
It is important to remember that on YouTube and Spotify, audio content competes with video content. According to Edison Research, YouTube has risen to the top as the most popular service used for podcast listening in the U.S.: 31% of weekly podcast listeners ages 13 and up choose YouTube as the service they use most to listen to podcasts, surpassing Spotify (27%) and Apple Podcasts (15%).
In the meantime, Spotify recently announced uninterrupted, ad-free video podcasts. It shared that nearly two-thirds of podcast listeners say they prefer podcasts with video.
In these datapoints you can see why Chaudry’s read on storytelling is partially right and partially wrong. It is right in the sense that it reflects how consumers still need and value human storytelling. NotebookLM may be producing video within the next few years but it may not matter: The challenges that OpenAI’s Sora has with rendering gymnastics suggests AI is not built for certain types of video storytelling.
But, it is wrong in the sense that consumer attention—the most valuable commodity in the Attention Economy—can still be captured by an AI-generated “performance” of an essay.
The Tension
There are no clear answers here.
The simplest takeaway from the above is that NotebookLM suggests AI will be less a storyteller and more a producer of faster, cheaper “performances”. It delivered the equivalent of a produced performance of my "story", and not an original story. Given that AI cannot produce compelling video of an essay analysis yet—if it ever will—the sense is that the “performances” AI will deliver may only be as good as the content humans feed it.
NotebookLM suggests AI will be less an original storyteller and more a producer of faster, cheaper “performances”. But, a storytelling "performance" is still storytelling, and therein lies the rub.
The platform has uncovered a grey area in "storytelling" that builders and creators had not seen before. The sense is that the debate around whether AI can replace human storytellers needs to rethought immediately. Because "performances" are storytelling, too: AI may not be able to write Shakespeare, but it will be able to perform it.