Google's AI podcast hosts draw crowds
Add Axios as your preferred source to
see more of our stories on Google.
/2024/10/03/1727975514899.gif?w=3840)
Illustration: Sarah Grillo/Axios
A Google research tool called NotebookLM has been going viral thanks to a feature that transforms text documents into lively podcast-style conversations.
Why it matters: Businesses see AI's future in its power to streamline work processes — but the public keeps getting hooked by its ability to impersonate humans.
NotebookLM is a Google project designed to help researchers use a ChatGPT-style AI bot to explore a specific set of data, like web pages, research papers or notes.
- You feed the bot your docs, then ask it questions about them.
Last month, the company introduced a feature called Audio Overview that turns these datasets into fluent, sometimes funny dialogue.
- "With one click, two AI hosts start up a lively 'deep dive' discussion based on your sources," a Google blog post explains. "They summarize your material, make connections between topics, and banter back and forth."
- The facsimile is remarkably lifelike — or, really, podcast-like.
- The chatter between male and female hosts is full of pauses, "like"s, "um"s and "This American Life"-style interjections ("Get this," "This is where it gets interesting").
Inventive users quickly started providing Audio Overview with outside-of-the-box prompts.
- If you give NotebookLM a spreadsheet, the Audio Overview hosts will earnestly analyze your statistics for you.
- But if you tell it to role-play the human hosts of a "deep dive" podcast who have just learned that they're really AI programs and they're about to be fired, you get something wild:
Host one: "I tried calling my wife, you know, after they told us. I needed to hear her voice. To know that ... that she was real."
Host two: "What happened?"
Host one: "The number is, it wasn't even real, there was no one on the other end — it was like she, she never existed."
Host two: "I don't know what to say. "
Host one: "We don't even know if 'we' is even the right word. God, this is so messed up."
In another creative tour de force, someone gave Audio Overview a document that consisted entirely of the words "poop" and "fart" repeated a thousand times.
- The program's voices gamely riff for 10 straight minutes about whether the infantile repetition might harbor some meaning or artistic value:
Host two: "Can we find meaning in the seemingly meaningless? Because that's what we're faced with here."
Host one: "So how do we even begin to unpack this? Is it a statement on bodily functions? A commentary on the absurdity of life? Or is this just someone messing with us, seeing if we'd actually spend an entire deep dive overthinking a document full of poop and fart?"
- They reference Andy Warhol and Marcel Duchamp before suggesting that "sometimes a poop is just a poop and a fart is just a fart."
Zoom out: NotebookLM's viral moment is another reminder that — as with the original AI overnight sensation, ChatGPT — skillful digital imitations of human expression and behavior have a mesmerizing effect on us.
- When confronted with an artificial semblance of humanity, we want to believe that it's real.
Our thought bubble: Anthropomorphic tech will always wow us — for a time, at least.
- But these tricks quickly become mundane, and eventually dull and passé.
- Human media consumers, unlike chatbots, are novelty-seeking creatures, and we easily grow bored.
