It started as a joke. “What if I created a podcast where I am talking to artificial intelligent bots and pretended to be interested in what they had to say?” I asked a friend of mine. 

That was a few months ago before I realized Google has an AI tool called NotebookLM that (mostly) lets you converse with two human-sounding hosts in real-time. Known as interactive mode, you can join the podcast and ask a question at any time. The bots will pause and recognize you as a “listener” and then comment on what you have to say.

I decided to test out NotebookLM to make a real podcast, launching it through Spotify for Creators and then promoting it on my social media channels. I learned a lot through the process, but the final result revealed quite a bit about what it takes to make a podcast. The one thing to avoid? Making an overly robotic sounding podcast without a soul. Here’s how I made it all work.

Recording in interactive mode

I didn’t realize it at first, but NotebookLM doesn’t actually record your interactions with the robot hosts. Because the app (available for mobile devices but it can also run in a browser) is in beta, there isn’t a lot of great documentation or even guidance on how to use it. Instead of relying on NotebookLM to record the interactive discussion, I used the iPhone Voice Memos app instead. While a true podcaster would scoff at the low quality, it was quick and easy.

To get started, you have to pick a source for the discussion. I decided to create “Origins with John Brandon” as a somewhat tongue-in-cheek podcast about how common things originated. My first episode was going to be about how bubble wrap came into existence, so I first found several articles about the topic and loaded them into NotebookLM. This is the training data the “hosts” use since they normally only talk to each other, but I didn’t want them to just start riffing on the topic. Instead, I interjected right away and introduced myself as the real host.

As you can imagine, this is not exactly how Google intended it to work. The bots acted surprised and even noted how I was not the host (mentioning me by name) and insisted they were the real human hosts. I asked a few questions about bubble wrap, but eventually the chat spiralled into a discussion about AI consciousness and awareness. I felt that was a much more interesting topic, so my podcast took on a more satirical flavor, similar to The Onion. In other words, while it was supposed to be about bubble wrap, it became more about AI.

I was happy with the results even if the recording quality was not outstanding. I imagined how it would have worked better if I would have recorded the live discussion with a separate digital recorder, but I wasn’t trying to win any podcast awards. I wanted to publish my first episode, and NoteBookLM had impressed me at least in terms of producing a coherent discussion. It was also cool how the bots responded to my comment in real-time.

Creating the assets

Of course, a podcast is more than just an audio file you publish. It also involves podcast art, bumper music, a voiceover, and other elements.

Once I had a recording for the first episode, I turned to Google Gemini to create a square tile I could use for the show. Every podcast has a tile that helps people identify the podcast in an app like Apple Podcasts. My prompt was simple: Create a square image for a podcast called Origins with John Brandon. I also provided an AI image of myself. On the first go around, Gemini created an exceptionally good tile, like something a professional might design. I didn’t like having my image associated with it, though, so I asked the bot to try again. The result looked a little cheesy but passable.

I was off to the races. I next used Suno to create the bumper music for my new podcast. This AI sound generation tool normally uses prompts like: Make a new song that sounds like Coldplay mixed with Nirvana. Instead, my prompt was more along the lines of asking for podcast bumper music with synths and drums to add some energy and excitement before we start talking.

Similar to my first attempt with Gemini, Suno cranked out some bumper music that sounded perfectly fine right away. I didn’t have to try any more prompts, but I did want to fine-tune the intro with a voiceover. I turned to Speechify, which can take any text you provide and make a voiceover. Then, I used ChatGPT to combine the bumper music with the voiceover as an overlay. Full disclosure, I did not know ChatGPT lets you make audio files.

Within minutes, I had intro music as a bumper with a voiceover. I even tweaked the crossfade between the music intro and when the voiceover starts and then had it fade out. I asked ChatGPT to add the podcast audio for my first episode and it generated one final audio file, which I uploaded to Spotify.

One thing to avoid

I was impressed with the results and how it all came together, but I would say I over-relied on AI for my new podcast. The tile looked slightly fake and soulless, especially since the tile image I used was generated by AI. The NotebookLM audio itself does have some more humanness and soul, mostly because I talk to the bots, but the episode ended up sounding like a human talking to chatbots, which is a recipe for disaster. People listen to podcasts to learn something new but also because the people talking are relatable, interesting, and real. At least it motivated me to make a real podcast without any bots involved on the same topic.

My main lesson was related to NotebookLM. It’s a good experimental tool and helped me figure out how to make and launch a podcast. Yet, it’s nothing but a lab experiment. The “hosts” repeat the same audio slop over and over, especially when I would join their conversation. Once you have heard them say “oh we have a listener who wants to comment” a few times, you start suspecting it’s all a computer-generated sham. They use the same voice ticks and quirky auditory mannerisms over and over again to the point where it is almost unlistenable and annoying.

In short, it’s boring. The podcast is not something I would want to keep online and mostly reveals that having bots talking to each other in this way is not going to hold anyone’s attention. It’s almost like watching two computers play a video game. The one thing to avoid with any podcast is relying so much on AI that it removes all humanity from the equation.

In the evolution of AI, we may need to decide if that is not a good outcome.