This weekend, I ran across an article from Engadget about how Google’s NotebookLM is now able to summarize YouTube videos and I had to try it out on some of my own videos. I’d tried NotebookLM before but it had been awhile since I hadn’t really found a use for it that ChatGPT couldn’t fill. This experience was different.

NotebookLM is an online language model tool that you can train with your own choice of materials including documents, pasted text, websites and now YouTube videos. These materials are grouped into notebooks within your profile and the language model will answer any questions for which it can find answers within the materials you’ve provided. It’s currently in the free, experimental stage and each notebook can hold up to 50 source documents.

NotebookLM’s addition of YouTube video analysis is a major step and I was very impressed how it was able to instantly access and add each video to the notebook. Although, considering Google’s ownership of YouTube, this easy integration probably shouldn’t be suprising. The only limitation I found was that it couldn’t do anything with the silent annotation-based videos I tried at first. It needs a transcript from voice audio.

Each notebook has a guide that enables the user to create various types of summaries from an FAQ to a study guide with a glossary, short answer and essay questions. This would be impressive enough and a very valuable resource for anyone wanting to research a subject. Then I noticed the “Audio Overview” feature in the top right-hand corner of the guide and decided to try it out. The result was pretty amazing.

The Audio Overview is actually a lively podcast-type conversation between two fictional hosts that goes into details about the material that you’ve uploaded. It’s the same, animated back-and-forth conversation that you would expect from any professional podcast and, aside from a glitch or two or the occasional mispronunciation, you would not know that it’s completely generated by A.I..

Once the overview is generated, you can download it from the notebook. Here’s the first one I generated on my C# control videos.

The promotional tone was a little over-the-top for the subject matter but I was still seriously impressed with the quality of the production. This was so far beyond anything I expected that it made NotebookLM’s primary functions seem a little mundane by comparison. It almost felt like Google buried the lead by including this as an extra feature.

Then I wondered what it would do with the Comeau Software homepage so I started a new notebook and fed it the main page. In just a few minutes, NotebookLM wrote better promotional copy than I could have imagined and generated a podcast from that would have cost serious money if I’d had an agency do it with live actors. I had to do a little bit of editing in Audacity to fix a mispronunciation of my name toward the end but, otherwise, it’s simply amazing to listen to.

I also tried out the feature on the pages for my Rogue C# project and upcoming SQLite book as well as the website I put together for a recent client project and every new result was equally impressive. That, in itself, is amazing for any new and innovative software feature – no major stumbles, no cases that it just couldn’t quite handle – it consistently performed, creating professional podcasts that, aside from being just a touch over-enthusiastic, could be immediately used for promotion.

How this would be used properly is another question. My first concern was that using something like this without an A.I. disclaimer would be inherently dishonest and a real risk to the credibility of whatever’s being promoted. A quick search of YouTube, however, shows that these Deep Dives with their standard host couple are becoming quickly recognizable as A.I. generated. Still, it’s dangerous to assume that everyone is as clued in to the latest trends as you are so, when I use them, I’ll be using a disclaimer like you now see on this site’s homepage.

Moving beyond their promotional use, these podcasts could be a useful tool for making summaries of information more accessible for people on the go or those with visual or reading challenges. On the other hand, they also risk becoming the modern equivalent of CliffsNotes, the overviews that our teachers always warned us not to rely on because they don’t convey all the details. It could be this battle is already being lost with the atrophy to our attention spans induced by social media and other modern technology but it still needs to be fought.

There’s also a subtle problem that, in trying to make the podcast more engaging, the virtual hosts claim actual life experiences in a way that blurs the line between the factual information being summarized and the fiction of an A.I. generated world. Here’s an example from the podcast that I generated for my article about the Technology Sabbath.

Host: I actually tried, like, a mini digital detox a little while ago and I was really surprised by how much more present, more engaged I felt in my just my everyday life.

Even when a listener is aware that they’re listening to a computer-generated fiction, the suspension of disbelief could act to give claims like this more weight and credibility than they deserve. On the other hand, there’s the question of how much an opinion from an unnamed host differs from the opinions that a person gets from fictional characters in movies and books that they’ve been consuming for years. While we hopefully realize that these characters are fictional, we become attached to them and their words influence us anyway.

For now, my ideal use of this new feature is to add accessibility to my longer articles. An audio overview can accompany the article, giving the reader a chance to hear an engaging, listener-friendly overview while still giving access to the original text for search. The A.I. is very good at taking complex subjects and making them more understandable to a general audience using the right metaphors and analogies. This might be a good compromise for writers like me as the online spaces continue to be crowded with more easily-digestible short videos and soundbytes.

Again, this is still an experimental feature and I don’t know what its future will look like. I can think of a couple of possible improvements such as variations on the hosts’ voices and the tone of the podcast. Maybe those will be included at a premium pricing level once NotebookLM leaves the experimental stage. Based on the other videos and articles I’ve seen about this feature and the demand it’s probably already placing on Google’s servers, I’m thinking that point might not be too far off.