Turning my last blog post into an AI-hosted podcast

Last week Google added a new capability to its AI-powered research assistant product, NotebookLM: the ability to generate an audio discussion between two AI ‘hosts’ based on any document you provide it with.

I tested it by pasting in the text of the blog post I wrote on Friday about the AI video generation race. Here’s the audio it created:

Observations:

• The presentation format owes a significant debt to podcasts: informal, conversational, imperfect.

• The ‘hosts’ are pretty plausible. They sound mostly human, with good variation in pacing and tone.

• There’s a fluid back and forth between the hosts (“exactly”, “yeah”, “right”, “really”, “definitely not”, “oh wow”, “you got it”, “that’s a great way to put it”, “I can imagine”, “oh, for sure”, “something like that”) and lots of questions (“is it?”, “so did they deliver?”, “how well did they handle the actual filmmaking part?”, “so which great leader should we start with?”, “any over-achievers?”)

• The use of filler words (“um”, “well”, “y’know”) adds to the verisimilitude, as do the affirmatory grunts (“uh-huh”). There’s even a simulated sneeze 32 seconds in.

• The discussion is fairly true to the structure and substance of my blog post.

• There’s still the odd hallucination (it made the leap from Roman leader to toga).

The option to have a machine read a document aloud to you has existed for many years. The ability to have a machine summarise a document before reading it to you is more recent. Having a machine create a passable facsimile of a human-hosted podcast from a single document is brand new.

It feels inevitable this - or something like it - will be used to churn out huge volumes of synthetic podcasts, as has already happened with books and music.

Less clear is the extent to which it will be adopted by learners looking to understand a topic in a format they find more accessible, especially once language support is broadened beyond English.

Enabling the listener to join the conversation feels like a likely future evolution, enabling them to ask questions and have them answered in real time.

Right now, I’m off to see if NotebookLM can turn the Renters’ Rights Bill into an engaging listen…