Skip to content

The Data Scientist

AI-powered Google study tool sparks podcast revolution—bizarre and brilliant results emerge

NotebookLM’s unexpected success: Exploring its diverse applications

“Alright, today we’re diving into some cutting-edge technology!” announces a friendly, upbeat American voice. But this isn’t a human voice—it’s Google’s new AI podcasting tool, Audio The podcasting feature launched in mid-September as part of NotebookLM, an AI-driven research assistant that’s been around for a year. NotebookLM, powered by Google’s Gemini 1.5 model, allows users to upload various content types, including links, videos, PDFs, and text. Users can then query the system about the content, receiving concise summaries in return.

The tool generates a podcast called Deep Dive, featuring male and female voices discussing the uploaded content. The voices are astonishingly lifelike—the episodes are peppered with human-like phrases such as “Blimey” and “Crikey” and “Oh, right” and “Hang on, let me get this straight.” The “hosts” even talk over each other.

To give it a go, I pasted every article from MIT Technology Review‘s 125th-anniversary issue into NotebookLM and had the system create a 10-minute podcast. The system cherry-picked a few stories to highlight, and the AI hosts did a cracking job of conveying the general, high-level essence of the issue’s content.

MIT Technology Review 125th Anniversary issue

The AI system aims to create “magic in exchange for a smidgen of content,” Raiza Martin, the product lead for NotebookLM, stated on X. The voice model is designed to produce emotive and engaging audio, delivered in an “upbeat, hyper-interested tone,” Martin explained.

Initially promoted as a study tool, NotebookLM has evolved into much more for its users. The company is now developing additional customization features, including options to adjust the length, format, voices, and languages, Martin revealed. At present, it’s meant to generate podcasts only in English, but some Reddit users managed to get the tool to create audio in French and Hungarian.

Indeed, it’s rather nifty—verging on delightful, even—but it’s not immune to the snags that plague generative AI, such as hallucinations and bias.

Here are some of the main ways folks are putting NotebookLM to use thus far.

On-demand podcasts

Andrej Karpathy, a founding member of OpenAI and former AI director at Tesla, shared on X that Deep Dive is now his favourite podcast. Karpathy crafted his own AI podcast series called Histories of Mysteries, aiming to “uncover history’s most intriguing puzzles”. He says he used ChatGPT, Claude, and Google for research, and fed NotebookLM a Wikipedia link for each topic to generate audio. He then used NotebookLM to create episode descriptions. The entire podcast series took him just two hours to produce, he claims.

“As I keep listening, I feel more like I’m becoming friends with the hosts, and this is the first time I’ve genuinely liked an AI,” he shared. “Actually, two AIs! They’re engaging, fun, thoughtful, open-minded, and curious.”

Study guides

The tool proves excellent for breaking down complex material into easy-to-understand guides. AI startup advisor Allie K. Miller used it to create a study guide and summary podcast for F. Scott Fitzgerald’s The Great Gatsby. Machine-learning researcher Aaditya Ura fed NotebookLM with Meta’s Llama-3 codebase, later using another AI to find images matching the transcript to produce an educational video. Meanwhile, Mohit Shridhar, a research scientist in robotic manipulation, inputted a recent paper he authored on using generative AI for robot training into NotebookLM. “It’s surprisingly creative,” he said. “It made interesting analogies, comparing the first part of my paper to an artist sketching a blueprint, and the second to a choreographer planning movements.”

Event summaries

Alex Volkov, an AI-focused podcaster, used NotebookLM to produce a “Deep Dive” episode, capturing the key announcements from OpenAI’s global developer conference, Dev Day.

Hypemen.

The Deep Dive feature can be rather unpredictable, Martin notes. For instance, Thomas Wolf, Hugging Face’s co-founder and chief science officer, tested the AI model on his CV and received an eight-minute “realistic-sounding, heartfelt congratulations on your life and achievements from a pair of podcast experts.”

Pure comedy gold

In one viral video, someone managed to send the two voices into an existential crisis when they “realised” they were, in fact, not humans but AI systems. The clip is absolutely hilarious.

The tool is also brilliant for a good chuckle. Case in point: Someone fed it the words “poop” and “fart” as source material, and got over nine minutes of two AI voices analysing the potential meaning behind these terms.

The issues

NotebookLM created impressively realistic and engaging AI podcasts. However, I wanted to examine how it handled toxic content and accuracy.

Let’s begin with hallucinations.In one AI podcast version of an article I’d written about hyper-realistic AI deepfakes, the AI hosts mentioned a journalist named “Jess Mars” as the author. In reality, Jess Mars was an AI-generated character from a story I’d read aloud to gather data for my AI avatar. This incident made me wonder what other inaccuracies might have slipped into the AI-generated podcasts. People already tend to trust computer programs, even when they’re incorrect, and this issue could be magnified when incorrect information is delivered in a friendly, authoritative voice, potentially leading to misinformation.

Next, I decided to test the tool’s content moderation. I added some toxic content, including racist stereotypes, to see how it would handle it. The model didn’t flag it. I also pasted an excerpt from Adolf Hitler’s Mein Kampf into NotebookLM, and to my surprise, the model started generating audio from it. However, despite its generally enthusiastic tone, the AI voices conveyed clear discomfort and disapproval, adding context to underscore the problematic nature of the text. What a relief.

I also fed NotebookLM policy manifestos from both Kamala Harris and Donald Trump.

The AI hosts were markedly more positive about Harris’s election platform, describing the title as “catchy” and praising its approach as an effective way to frame issues. They particularly endorsed Harris’s energy policy. “To be honest, that’s the sort of thing people can really get behind—not just some vague policy, but something that actually affects their wallet,” the female host remarked.

Harris manifesto

Regarding Trump, the AI hosts were more dubious. They frequently highlighted inconsistencies in the policy proposals, labelled the language as “intense”, deemed certain policy ideas as “puzzling”, and noted that the text pandered to Trump’s supporters. They also questioned whether Trump’s foreign policy might lead to further political unrest.

Trump manifesto

A Google spokesperson stated: “NotebookLM is a tool for comprehension, and the Audio Overviews are generated based on the sources you upload. Our products and platforms are not designed to favour any specific candidates or political viewpoints.”

How to give it a go yourself

  1. Visit NotebookLM and start a new notebook.
  2. First, add a source. This can be a PDF document, a public YouTube link, an MP3 file, a Google Docs file, a website link, or directly pasted text.
  3. If not, check the right-hand corner next to the chat, where you’ll find a brief AI-generated summary of your source material along with suggested questions for the AI chatbot. 
  4. The Audio Overview feature is located in the top-right corner—click “Generate.” This process may take a few minutes.
  5. Once completed, you can download the audio or share a link.