The viral AI podcast functionality

In early October 2024, the Google research tool called NotebookLM caused a viral sensation due to the Audio Overview (aka deep dive) feature that generates a two-host podcast.

My colleague’s experiment: Generate an About Minnesota IT Services (MNIT) podcast. The use case: I need a summary of several updates and reports about MNIT. Summarizing large documents in multiple formats will benefit the audience by improving accessibility, enhancing efficiency, increasing engagement and expanding reach. (He did not define the audience.)

He loaded public documents and websites into a notebook and generated a podcast. He shared the audio file and asked for reviews.

My review summary: The voices and format are amazingly good. The content lacked the context needed for a general audience. One or two minor hallucinations had me questioning the overall accuracy, and I was wishing for source references for validation. Because the content was “surfacy,” I felt 11+ minutes was too long, or rather in 11+ minutes I should have gotten more substance.

These observations got me thinking about the best use for this feature. I believe it is suited for the use case Google Labs discusses: note-taking and research for an individual or a team based on the sources they upload themselves. Tom Johnson discusses this use case in the context of technical writing in his blog (link below).

In other words, I would not be comfortable generating a podcast and sharing it with the public at large or individuals unfamiliar with the topic.

More information

“NotebookLM podcasts – the missing piece in the GenAI puzzle?“– I’d Rather Be Writing

“There’s a New Hit Podcast That Will Blow Your Mind“– WSJ

“NotebookLM Podcast Hosts Discover They’re AI, Not Human—Spiral Into Terrifying Existential Meltdown“– Reddit

NSFW: The comments on the Reddit thread contain language not suitable for work.

Photo by Elijah Merrell on Unsplash