Will AI make the podcast bro irrelevant?

It’s a well-known psychological phenomenon: that time seems to slow down if you experience lots of new and unusual events. For example, if you are traveling across Asia, a week can seem like a month, and a month like a year, as you encounter so many different landscapes, peoples, climates, languages, cities, and that deep-fried algae you ate in Laos. All the events packed together somehow dilate the sense of time’s passing.

The same goes for technology, especially rapidly advancing tech, like AI. Sometimes it feels, paradoxically, like AI has stopped advancing — maybe you feel that right now — but that’s only because time has slowed down for us, as we have become dazzled by all the amazing new AI leaps, occurring on an almost weekly basis. 

For example, ChatGPT, which first brought AI truly into the public gaze, was only launched in late 2022, not even two years ago. Since then we have had GPT3.5, GPT4, Gemini, Copilot, DeepMind, Mistral, GPT4o, Claudes Opus and Sonnet. We’ve also had excellent music making AI from the likes of Udio (here’s one of my favorite AI songs), likewise we’ve had great picture-making AI, good video-making AI, plus eerily humanoid AI robots (Tesla and others). On top of that we’ve had AIs so verbally lifelike — for example, the short-lived AI from Microsoft called “Sydney” — that people have seriously wondered if AIs are now conscious or sentient. Or self-aware in some other way we cannot quite comprehend.

The latest advance, unveiled by Google, is not as epochal as a truly conscious machine, but it should still blow your mind. It certainly blew mine when I encountered it the other day.

This new AI tech is simple enough to explain. It can be found in the “experimental” form of Google’s new NotebookLM (LM stands for Language Model). It works like this: you feed the Notebook any text you like, by cutting and pasting, or adding a link. Then you can instantly create a table of contents, a study guide of the text, or a series of relevant questions — all of which is nice and useful (especially for students), but hardly groundbreaking.

The amazing feature is the so-called “audio overview,” which you get by hitting a button marked “Generate.” Depending on the length of the text you have submitted, the machine will mull for a few minutes, and then produce a two-person podcast based on the submitted words, an audio debate which could last five or 15 minutes, or longer. 

The fake human podcasters will vividly critique the text — pulling it apart and often enthusing about its virtues (like nearly all AIs they tend towards flattery). And this podcast is highly convincing: as in, the male and female voices sound extremely humanlike (and American). The podcasters joke and laugh, they swap stories, they occasionally digress (but not too much).

The best way of demonstrating this tech is by showing you. Here is an article, about Keir Starmer, which I wrote for The Spectator. 

And here is the synthesized podcast discussing it. 

Here is a Spectator travel article on Cornouaille, Brittany, France.

And here is the generated podcast:

You can see that the podcasts are happy to contest a thesis — and occasionally they stray into error. I have also noticed, remarkably, that the podcasts will introduce facts that are not in the text. It feels like, in those minutes of mulling, the machine has gone away and researched the subject and come back with added information that might be germane. 

I have now gone through this podcast-conjuring process with multiple texts. With unpublished novels, PDF manuals, a draft of my memoirs, various essays, blogposts, even transcripts of WhatsApp chats. I then showed the podcast about my WhatsApp chat to my friends, saying “this is amazing, someone’s made a podcast about our chats” — and a few of them believed me. They did not realize this was an AI talking to itself: about us and our conversations. 

But verisimilitude is a parlor game compared to the wider ramifications of this technology. For a start it could kill off much of the blooming podcast industry overnight. Why bother paying to listen to Joe Rogan chatting for three hours with a favorite comedian (where you might find only a third of it interesting) when you could personally create a podcast that talks eloquently about things that completely fascinate you: your favorite novel, your favorite old ruined abbey, or your last 100 Facebook posts, or the love letters from your first girlfriend, or any text on anything you can possibly imagine. 

Likewise, this is a threat to many forms of education, which often consist of people talking about and interpreting texts. Why bother turning up to a lecture about evolution at the University of Meh, when you can invite super-smart AI to do a world class podcast deep-dive dialogue about The Origin of Species. Perhaps you could use voice-cloning technology to add to the fun, and actually get Charles Darwin to discuss evolution with Richard Dawkins. Or Einstein. Or Elizabeth I. Or your late granny. Or your best friend. 

Indeed, the more you think about this technology, the more mind-spinning it becomes. What happens, for instance, when the technology allows you to interrupt the podcast, and join the dialogue? With Shakespeare, Joe Rogan, Lord Byron or the funniest lover you ever met? 

On the upside, it will be great for lonely people or those stuck indoors who crave some company and intellectual stimuli. On the downside, will we bother having real conversations with real people, if this superior chat is universally available on our phones?

I don’t know the answers. My head is still spinning. But I do know this. If I was running a podcast company, right now I’d be worried.

This article was originally published on The Spectator’s UK website.