Google just added a new customization tool for the viral AI podcasts in its NotebookLM software. I got early access and tested it out using Franz Kafka’s The Metamorphosis as the source material, spending a few hours generating podcasts about the seminal novella—some of them more unhinged than others.
Released by Google Labs in 2023 as an experimental, AI-focused writing tool, NotebookLM has been enjoying a resurgence in user interest since early September, when the developers added an option to generate podcast-like conversations between two AI voices—one male-sounding and one female-sounding—from uploaded documents. While these audio “deep dives” can be used for studying and productivity, many of the viral clips online focused on the entertainment factor of asking robot hosts to discuss bizarre or highly personal source documents, like a LinkedIn profile.
Raiza Martin, who leads the NotebookLM team inside of Google Labs, is pumped to give users more control over the content of these synthetic podcasts. “It's the number one feature we've heard people request,” she says. “They want to provide a little bit of feedback as to what the deep dive focuses on.” According to Martin, this is the first update of many coming down the pipeline.
Nearing the one year anniversary of its full launch, NotebookLM is also dropping the “experimental” tag—a sign it’s not headed towards the perpetual Google graveyard of abandoned software, or at least for now. Martin says this label was removed since the team hit internal milestones for overall quality, user retention, and interface standards. She also says users can now expect a higher level of stability from the software.
How to Customize the AI Podcasts
To make an AI podcast using NotebookLM, open up the Google Labs website and start a New Notebook. Then, add any source documents you would like to be used for the audio output. These can be anything from files on your computer to YouTube links.
Next, when you click on the Notebook guide, you’ll now see the option to generate a deep dive as well as the option to customize it first. Choose Customize and add your prompt for how you’d like the AI podcast to come out. The software suggests that you consider what sections of the sources you’d like highlighted, larger topics you want further explored, or different intended audiences who you want the message to reach.
One tip Martin shares for trying out the new feature is to generate the Audio Overview without changes, and while you’re listening to this first iteration, write down any burning questions you have or topics you wish it expanded on. Afterwards, use these notes as a launching pad to create your prompts for NotebookLM and regenerate that AI podcast with your interests in mind.
My First Impressions
I uploaded an 80-page file of Kafka’s famous work of existential literature—in it, the main character wakes up one morning to find that he has turned into a gigantic bug—to see how the customization will work for NotebookLM users. The first Audio Overview it generated, sans prompt customization, was a solid, albeit broad overview of what happens in the novella, as well as some discussion of its key themes. Nothing groundbreaking, but decent.
Thinking like a nerdy college English major, which I definitely was, my first prompt adjustment was for the podcast discussion to focus more on themes of alienation and overbearing bureaucracy found in the book. With the extra nudge, this output from NotebookLM did an admirable job of zeroing in on these motifs and generating a discussion that sounded similar to what I’ve heard before in college classrooms. It was a bit meandering, but totally listenable.
Next, I asked the tool to focus on a few specific pages from the source document. This result was more underwhelming than the rest of the tests, since it sounded fairly similar to the AI podcast generated without a custom prompt. To be fair, if I had uploaded more source material to NotebookLM and the AI had more to work with, rather than just a couple pages from a PDF, the instructions to concentrate on specific aspects may have produced more satisfactory results.
By far the most interesting way to tweak the AI podcasts, for me, was requesting the synthetic hosts focus on varied intended audiences. When asked to explain the novella to a recent college grad starting their first office job, the hosts astutely discussed what it’s like to go through major life changes. Though, I almost spit out my coffee from laughing at NotebookLM’s attempt to translate Kafka’s prose for a fierce group of drag queens. “Honey, let’s dish about the isolation,” says the male-sounding AI voice. “Because Kafka does not hold back.”
While some of the slang used by the hosts to communicate with the drag queens was nonsensical and a bit cringe, I was impressed to see how the output was tailored to highlight aspects of the novella queer people might empathize with, like Kafka’s focus on otherness and family tensions. Yes, Google’s NotebookLM might flatten the specifics of a big document or get some details mixed up, but being able to generate more personalized podcasts from disparate sources truly does feel like a transformation—and luckily nothing like turning into a giant bug.
