Here we go again—another generative AI service churning out plausible-sounding, but misleading content. This time it’s Google’s NotebookLM, branded as “your personalized AI research assistant powered by Google’s most capable model, Gemini 1.5 Pro”. It now includes Audio Overview, which creates audio discussions between “two AI hosts” who supposedly engage in a “lively deep dive” based on your sources. But Google’s Audio Overview tool in NotebookLM generates bogus output.
As I’ll explain below, when I tested Audio Overview on last week’s blog post, it generated a five-minute “conversation” that misrepresented the content. It introduced contradictory information, and fabricated topics that didn’t exist in the original post.
Why Audio Overview is concerning
The real danger of this tool lies in its ability to produce convincing, natural-sounding, audio conversations between two computer-generated voices. In today’s digital landscape, most people are becoming more skeptical about what they read. But an audio recording of two people enthusiastically summarizing content sounds far more trustworthy, even if it’s completely AI-generated.
How I tested NotebookLM’s Audio Overview
Testing the Audio Overview feature was quick. I created a new Notebook, uploaded the text of my recent blog post titled “How the Responsibilities of Conferences Mirror Those of Media Platforms,” and waited for the tool to generate its insights.
As you’d expect from LLMs these days, NotebookLM provided a good written summary of the post:
“The sources compare the responsibilities of news organizations to the responsibilities of conferences, arguing that both should prioritize their users and be transparent, open, and reliable. The author draws parallels between Jeff Jarvis’s principles for media platforms and his own principles for conference design, which emphasize participant-driven, collaborative, and ethical practices. Both sources advocate for a shift away from outdated, broadcast models and towards user-centric approaches that foster trust and empower participants.”
So far, so good. But then, I clicked on the Audio Overview feature, and that’s where things took a turn. I’ve added the audio and a transcript below.
Listen to at least some of the audio to understand how eerily convincing such an AI-generated “conversation” can be. You’ll hear what sounds like two people constantly interrupting each other in their enthusiasm to share their 100% positive take on the post.
Audio track
Transcript
Woman: Ever feel like you’re stuck in a conference rut? You go, you sit, you listen, maybe snag a pen or two.
Man: Yeah.
Woman: But something’s missing. Like it doesn’t really change anything back at your desk.
Man: Right.
Woman: That’s what’s got me so fired up about this deep dive. We’re flipping the script on conferences, how to make them crazy, valuable, engaging, all that.
Man: Okay.
Woman: Get this. Our source. It’s this blog post comparing, get this, designing online platforms like the Big Dogs, Google, Twitter, to designing amazing conference experiences.
Man: Totally.
Woman: I know, right? Total apples and oranges, you’d think?
Man: Totally.
Woman: But stick with me. This article makes a killer case. Borrow some key things from those platforms, that user first mindset.
Man: Okay.
Woman: And boom, revolution in conference land.
Man: See, that’s what I’m talking about, because what’s always bugged me is how we treat conferences.
Woman: Yeah.
Man: It’s always been the same old, experts on stage, everyone else just sits there, soaks it up.
Woman: Passive.
Man: Exactly. But the platforms we use every day, they’re all about putting you in control. Choices, shaping the experience, big difference.
Woman: Huge! And this Jarvis guy, he wrote the article, nails it. He lays out these parallels, like how platforms are all about users controlling their own data, right? And he says, hand over control to your users, give them power, design and flexibility. Imagine that, but for conferences, forget the fixed schedule, you choose what sessions actually speak to you, your own path.
Man: It’s not just showing up anymore, right? We’re drowning in information everywhere else. I want to grab what matters, meet the right people, walk out with, like, usable knowledge, you know?
Woman: Totally. And that’s where this openness thing comes in, Jarvis talked about. This whole data portability thing really stuck with me. In the online world, it’s like, taking your stuff with you, contacts, files, whatever, when you switch platforms, no biggie.
Man: Right, right.
Woman: So how do we steal that? For conferences, make them really pay off for people, you know.
Man: It’s that feeling walking out and it’s not just a blur. Like what was that keynote again? No more business cards. You’ll never actually…
Woman: Tell me about it. The black hole of business cards.
Man: Gone. You leave with stuff you can use.
Woman: Okay, I’m with you. But how? How do we take data portability? Make it real for conferences.
Man: Picture this, conference app, right? Not just the schedule, but as you’re there listening, you’re highlighting stuff, snapping pics of slides.
Woman: Instead of scribbling like mad while they talk.
Man: Exactly. Your own notes, all in the app. Then, boom, conference ends, it bundles it all up, highlights notes, even like who you need to connect with, your own little takeaway package ready to go. That’s data portability.
Woman: No more digging through that stack of notes a month later trying to remember.
Man: Exactly. You get max value. You’re not just consuming, you’re curating. Putting it to work.
Woman: It’s so obvious when you say it like that. But honestly, so many conferences I walk out. Yeah, I learned stuff. But what am I going to do?
Man: Missed opportunity, right? Yeah. For everyone. You’re putting in the time, the money, you deserve to get something back.
Woman: For sure. Okay. So say we buy into this whole user first thing. Where do we even start? How do organizers make it happen? That’s a whole new way of looking at it, right? Not just an audience in the dark, but they’re part of it, shaping the whole thing.
Man: And that starts way before anyone even shows up. What if, like, before the conference, bam, organizers send out surveys?
Woman: Okay.
Man: Not just “How’d you hear about us?”
Woman: Yeah.
Man: But what are you hoping to get out of this? What are your goals even? How do you learn best?
Woman: I love that. It’s like actually caring if people get something out of it. Like, say it’s about the future of work or something. Wouldn’t you be way more into it if you had a say in what they even talked about?
Man: Totally. And picture this. What if there were, like, online forums, but before the conference?
Woman: Okay.
Man: People connecting, saying what they’re hoping for, starting the conversation early. By the time they get there, they already know people.
Woman: No more awkward icebreakers. You’re already halfway there.
Man: Exactly. And then the conference itself. Ditch those long, boring presentations, right? Shorter, punchier, more hands on stuff, workshops, solving problems together.
Woman: Instead of just sitting there, you’re doing something, working with people, trying out new ideas, so much better.
Man: And feedback. Not just those forms you forget on the way out.
Woman: Yeah.
Man: But all the way through, telling them what’s working, what’s not. They actually listen and make it better.
Woman: So you’re always tweaking, making it better as you go, based on what people actually need.
Man: Exactly. People feel like they’re part of it. Like their voice matters. That’s community, right there.
Woman: So, bottom line, maybe it’s time we stop just accepting the same old conference, you know?
Man: It’s time to expect more.
Woman: We can be those users demanding better, and who knows? Maybe change the whole game. That’s what I’m taking away from this.
Man: Me too. It’s about potential. Conferences could be so much more than we settle for.
Woman: Totally. Thanks for diving deep with us. Hope this got the wheels turning about what these things could be. See you next time.
How NotebookLM’s Audio Overview went wrong
Everything I wrote is wonderful!
My post about how the responsibilities of conferences mirror those of media platforms was an opinion piece. NotebookLM’s Audio Overview turned it into a PR flack’s dream.
Can NotebookLM be critical? Does it ever respond: “I can’t summarize this post intelligently or accurately?”
Somehow, I doubt it.
Despite sounding fluent and convincing, the generated “conversation” was riddled with inaccuracies. Here are some major issues:
Who wrote the blog post?
Hey, Audio Overview, I wrote the blog post, not Jeff Jarvis!
Misrepresentation of the responsibilities of platforms vs. conferences
The “woman” in the conversation says the blog post compares designing online platforms like Google and Twitter to designing conferences. This isn’t accurate. My post compares the responsibilities of media platforms and conferences, not their designs.
Exaggeration of user control
The conversation suggests that platforms like Google or Twitter are “all about putting you in control,” implying complete user freedom. However, my post and Jeff Jarvis’s quote are more nuanced. While Jarvis advocates giving users power and flexibility, there’s no suggestion that media platforms fully hand over control to users.
Invention of new topics: data portability, apps, and live feedback
The audio invents topics I never mentioned, such as data portability (taking contacts and files when switching platforms), a high-tech conference app where participants take notes and pictures, and vague musings about live feedback during the event. While these are interesting ideas, they have no basis in the original post, where I discuss the portability of the peer conference process, not technology.
Fabrication of pre-conference engagement
The “man” talks about sending surveys to conference attendees before the event, to ask what they want to learn. While this could be a useful idea, I didn’t mention anything about pre-conference surveys in the blog post. The blog discusses uncovering participants’ wants and needs during the event, not beforehand.
Conclusions
Audio Overview in Google’s NotebookLM generates bogus output. The tool transformed my blog post into an overly simplified, inaccurate PR pitch for conference engagement. What concerns me most is that it does this convincingly by aping the conversations humans have used to communicate and connect for over a hundred thousand years. Consequently, users may not question the validity of what is said because the format is familiar and comfortable.
But wait, there’s more! In my next post, I’ll explore how the Audio Overview service handles factually incorrect content. Stay tuned for my next test!