I found that the audio overviews fell short in a key way, but were still impressive.
Earlier this month, Google superpowered its AI research and writing tool, NotebookLM, with an experimental Audio Overview feature that transforms any collection of sources into a podcast discussion between two AI hosts. Google advertises the AI discussion as downloadable, engaging, and a tool for auditory learners.
But it also sounds like a podcast between two human beings, with masterful pacing, tone, and delivery -- after trying it, I think the tool can inspire human podcasters.
Related: I Tried 3 AI Headshot Generators and There Was a Clear Winner
Google Vice President Josh Woodward previewed the feature at Google's I/O event in May, stating "You can give [NotebookLM] lots of information in any format and it can be transformed in a way that's personalized and interactive for you."
Intrigued, I put the feature to the test.
Google acknowledges that at this stage, NotebookLM's AI podcasts are limited to English audio and they "sometimes introduce inaccuracies." They aren't interactive yet, and can't be interrupted.
I first went to NotebookLM and created a new notebook. I then started adding sources, focusing on broad, dense, publicly available materials -- I found textbooks in the public domain about everything from music theory to calculus.
Google allows up to 50 sources per Notebook. I could upload PDF, .txt, or Markdown files from my laptop or choose sources through Google Docs, Google Slides, a website link, and copied and pasted text.
I relied mainly on website links and PDFs and chose eight sources, each in a different discipline with hundreds of pages. Leonardo Da Vinci's Thoughts on Art and Life, Silvanus Thompson's Calculus Made Easy, Robert Hutchinson's open-license college textbook Music Theory for the 21st-Century Classroom, James B. Conant's Organic Synthesis, Edgar Thurston's Omens and Superstitions of Southern India, and Jacob Joshua Levison's Studies of Trees.
I also added some philosophy: Beyond Good and Evil and The Republic, both through Project Gutenberg. I wanted to see how the AI podcast would draw connections between the texts and I wondered how the AI would process them into a digestible audio.
Related: Can ChatGPT Help Start a Business? I Tried the Latest Version, GPT-4o, to Find Out.
After uploading the sources, I clicked "Notebook guide," went to "Audio Overview," and clicked the "Generate" button next to "Deep dive conversation."
After 3 minutes and 57 seconds, the audio overview was complete. It was 8 minutes and 53 seconds long and featured two voices, one female and one male.
"Okay so you think lucky socks are cool," the female voice said, starting the podcast. "Well get ready because today we're going way beyond that."
"Way beyond that," the male voice added.
The audio presented was engaging, with a constant back-and-forth between the two hosts, but it only focused on the Edgar Thurston text and neglected everything else. That text wasn't the first source on the list and I didn't single it out in any way -- the AI seems to have picked it at random. The way the AI hosts presented the text was impressive and conversational, but I was expecting something more comprehensive that encompassed all of the sources and drew connections between them.
When I deleted the Thurston text and tried to reload the audio overview, it immediately gave me the podcast it had already generated, even though the audio no longer aligned with the sources.
Related: New Google Report Reveals the Hidden Cost of AI
I was able to download the podcast easily, and even if the audio focused on just one source, it went deeper into the source and made it understandable. For students, I can see it being a useful study aid or a way to understand more densely written texts. For anyone trying to understand dense research or books, the tool could deliver audio conversational enough for a morning commute.
Having a personalized podcast is a powerful tool -- after creating one myself, I'd say it's a strong use case for AI.