Is It NPR or AI? Google Turns Your Notes Into Incredibly Real-Sounding Podcasts

Listen to this. If we didn’t tell you it was generated (in a few minutes!), we doubt you’d know it was entirely done by AI.

This stuff is getting wild.

The latest entry in the race to make human creativity obsolete is Google’s turbocharged AI-powered “NotebookLM,” which offers a feature that nobody knew they wanted: It can turn your notes and documents into a podcast so real-sounding that the unsuspecting might confuse it with “Morning Edition.”

The “Audio Overview” upgrade, announced on September 11, transforms uploaded content into lively, intelligently modulated discussions between two AI hosts.

The AI hosts dive deep, summarizing content and drawing connections between topics, offering users a fresh way to absorb information.

That means that for now at least, an aspiring media magnate could, for instance, paste in stories about Tuesday’s pager attack in Lebanon from, say, the New York Times, Washington Post, The Guardian, Al Jazeera, Haaretz, CNN and BBC, and in a few minutes make a 10-minute podcast that rivals anything out there.

Since this is considered interpretive rather than outright content theft, one could even create a daily podcast from the world’s best sources, and sell ads against it. Here’s what that would sound like.

But podcasts are just the tip of the iceberg. NotebookLM has evolved into a powerhouse for processing and contextualizing massive amounts of data. It now handles over 50 different sources simultaneously, synthesizing information for personalized insights. This puts it leagues ahead of competitors that rely on basic keyword searches.

The upgrade leverages Google’s Gemini 1.5 model, which processes over 1 million tokens. This means richer, more comprehensive contextual insights. For researchers, students, and professionals drowning in data, NotebookLM offers a lifeline, generating study guides, briefing documents, and even sparking new ideas.

Since its 2023 debut, NotebookLM has gone global and expanded its toolkit. It now supports Google Slides and web URLs, demonstrating enhanced multimodal capabilities. NotebookLM’s ability to “ground” its AI in user-specific documents appears different from what other similar chatbots and AI tools do by relying on keywords and local search for their retrieval tasks.

The Audio Overview feature, while groundbreaking, isn’t without limitations. It’s currently English-only and can introduce inaccuracies. We tested the model with Spanish sources, and it produced a coherent podcast… in English. In our tests, the model didn’t generate any factually fake information, however, it did slip in some contextual things that could be spotted by wary listeners.

How to make your own podcasts

Generating your own podcasts with the tool is actually pretty easy.

Just go to the official NotebookLM site and click on the “Try NotebookLM” button at the top right.

Once you are in, you can either play around with some pre-loaded examples—interact with them, listen to those podcasts, and get familiar with the tool—or you can generate your own podcasts. To do so, click on the tab that says “New Notebook”

Once in, you’ll find a pretty intuitive UI. The first thing you need to do is add the sources that NotebookLM will analyze to either interact with you via chat, or generate the podcast. Should you go the chat route, the AI will be all text, helping you understand your information and present it better. It won’t mess with the podcast’s format.

You can link your Google Drive account and upload your personal files from there, add URLs from any site you want (it will only understand the text that is shown in that specific URL), or directly paste in the text you want it to analyze.

These options are not mutually exclusive. You can mix them and add up to 50 different sources

You can work from the popup window, or add sources manually by closing the popup and clicking the plus button in the left panel. If there is a specific source you don’t want the model to exclude, you can uncheck it.

If you want to see a brief summary of a specific source, just click on it and you’ll see an AI-generated paragraph with the most important information, key topics, and the source’s whole text below it.

If you don’t want to make a podcast from your notes, but are instead, say, researching the topic at hand, you can now query the notebook you created. Imagine it’s a friend that knows every single thing there is to know about the information that you provided. Go to the little text box in the lower part of the screen and ask your questions.

NotebookLM will give you a very detailed reply. Its answers will be relevant and likely beat any information that a normal AI chatbot would provide. Note that like any other AI model, it will sometimes “hallucinate” and provide wrong info—so if you are planning to publish or otherwise disseminate the information, then you’ll need to double-check it.

This is very powerful already, especially for users looking for ways to analyze huge amounts of data. For example, you can upload a lot of information on a company, and then ask questions to help you decide whether it is wise or not to invest in it. Or you can upload a bunch of research papers, and then ask for relevant information about previous studies on a subject.

But the fun really begins when you ask NotebookLM to generate a podcast about your notes. If you were chatting with the model, then simply switch to the “Notebook Guide” button on the lower right part of the screen and click on the “Load Conversation” button in the Audio Overview section

Then all you have to do is wait. In a few minutes, NotebookLM will generate a 10-minute podcast with two AI speakers discussing all of the key topics from your sources.

We put NotebookLM through its paces using a real-world test case. Taking different sources from the extensive coverage of Donald Trump’s interview Monday with Rug Radio—Decrypt‘s sister company—we fed them into the system. And the results were actually pretty good.

In this case, NotebookLM churned out fluent content that didn’t veer into hallucination territory. It served up solid insights, drawing connections we hadn’t explicitly made. Internally, our team was impressed with the output, and some couldn’t tell it apart from human-written analysis.

NotebookLM cannot generate more than 10 minutes of audio, however, so human-generated, long-form content is safe—for now.

Edited by Josh Quittner and Andrew Hayward