NotebookLM makes shockingly good audio

Issue #45

On today’s quest:

— NotebookLM makes shockingly good audio
— Perplexity will soon have ads
— Can you spot the bot-written reviews?
— Can anyone spot AI photos?

NotebookLM makes shockingly good audio

It’s been a long time since I’ve been shocked by the quality of an AI product, but it just happened when I clicked the “Audio Overview” button in Google’s NotebookLM and got something that sounded exactly like a podcast with two hosts talking, and at first glance* it rivaled most of the podcasts I regularly hear in my feed.

What is NotebookLM?

Let me back up. NotebookLM is a tool for organizing your research. The descriptions about it can be confusing, but a friend has been after me to use it for a while, and I finally tried it.

In brief, you create a project and add your research materials such as PDFs, web links, and notes you cut and paste. Then, you can use the AI to do things with those materials — create a list of all the people mentioned, put events in chronological order, and ask it to list all the alternative final lines in old versions of a nursery rhyme, for example. I’m finding that I like it more than I thought I would.

What are Audio Overviews?

I wasn’t testing it with anything related to podcasting in mind, but it turns out that, coincidentally, Google had launched “Audio Overviews” in NotebookLM a few days earlier, so when I saw the button, I clicked it. And I got audio with a masculine and a feminine voice having what sounds like a completely natural, engaging conversation about the topic I was researching.

After pulling my jaw up off the floor, I called my husband over and said, “You have to listen to this,” and also felt like I needed to add, “These are not real people.”

You can hear an example about halfway down the page on Google’s blog post about the launch, and it is representative of what I heard.

A new high for AI audio

As a podcaster, I have been keeping a close eye on AI audio. I’ve listened to a lot of AI-generated podcasts and have found them to be lacking, sometimes cringey, and just plain uninspiring. This audio is nothing like those. It blows past the threshold of listenability.

My gut reaction was to actually tell no one. I don’t want people to know about this. I don’t want to have to compete with AI-generated podcasts. But of course, that’s ridiculous. I can’t stop it from existing, and the whole reason I started this newsletter is because I think it’s important to know what’s happening and what’s coming with AI.

But with some downsides (for now)

The Audio Overviews do have a bunch of downsides. You can’t edit the audio, so if you find a problem, there doesn’t seem to be a way to fix it. You also can’t select the voices, so if you don’t like the defaults, you’re stuck.

Jeff Jarvis played with it more than I did and talked about his experience in his AI Inside podcast (the NotebookLM section starts around 24 minutes). He generated multiple Audio Overviews for the same content and found the quality of the content to be uneven.

I expect most of these problems to be fixed, but even if they aren’t, right now people can definitely make things that sound like high-quality podcasts with interesting, mostly accurate content on any topic in minutes.

What’s a human to do?

We’re about to be flooded with good AI audio content, and that content is only going to get better.

So what can podcasters do so we’re not buried by AI? The same thing I’ve been telling writers and editors: provide the confidence that we’ve checked the accuracy of our work and focus on human connection.

A great AI podcast will definitely get listeners, but I’ve also listened to podcasts with bad audio because I liked the hosts or the particular kind of content they made. I’ve occasionally listened to podcasts on topics I barely care about because I like the hosts.**

And I’ve stopped listening to podcasts whose content I did like because I didn’t like a host’s voice or something else about them. I support Patreons for podcasters whose podcasts I never listen to because I like the work they do for other people, and I want their podcasts to exist.

Yes, the online content world is about to get even more competitive, but humans also don’t consume content (or hire freelancers) for completely rational reasons. The best thing we can do is focus on quality and our humanity.

Perplexity will have ads soon

I know many of you like using Perplexity for AI searching. The company just announced it will be incorporating ads by the end of the year.

Can you spot bot-written reviews?

It was harder than I expected to spot bot-written hotel reviews on this 15-question quiz! I got only nine right, which was painfully humbling. (h/t Nancy Friedman)

I misclassified three bot-written reviews as coming from a human, and one human-written review as coming from a bot. I lost three more points by being unsure and undecided (ranking two reviews with a lot of emoji as “maybe bot” instead of “definitely bot” and clicking “undecided” on one review, which lost me two points).

IF YOU WANT TO TAKE THE QUIZ, do it before you read the next section with my analysis.

This was my reasoning for my choices:

• Emoji. I'm very suspicious that anything with an emoji is written by a bot given its tendency to include them. (I should have embraced this harder. It was correct in every case.)

• Writing errors. I'm almost certain that anything with a major English language error is written by a human since I've never seen a bot make an "its/it's" error or forget to put a space between words. (This was correct in every case.)

• Long reviews. I tend to think things that are long, overly effusive, or use “fancy” words were written by bots. (I got fooled by a short bot-written review, and after seeing the results, I should have paid more attention to the fancy words including “awaited” and “showcased.”)

• Details. If something has too much detail (like the names of the writer's children), it feels like a bot. (This was correct.)

• Positivity. I'm more likely to think a review that includes something negative is written by a human since I expect bots to be used to pump up a hotel's reputation. (I marked one especially effusive review with one awkward phrase as written by a bot, but it was actually from a human).

Can anyone spot the AI photos?

Much like my problem with the restaurant reviews, a photographer who has been concerned about manipulated photos for decades recently described being unable to accurately identify altered photos. In the Today Explained podcast, he talked about the effect AI photos are having on our perception of reality (and made an argument for why it’s a problem to even fix family photos so everyone’s eyes are open***).

What is AI sidequest?

Using AI isn’t my main job, and it probably isn’t yours either. I’m Mignon Fogarty, and Grammar Girl is my main gig, but I haven’t seen a technology this transformative since the development of the internet, and I want to learn about it. I bet you do too.

So here we are! Sidequesting together.

If you like the newsletter, please share it with a friend.

^{* What’s the audio equivalent of “glance”?}

^{** I have never been a Car Talk listener, but I hear people say this about that show a lot — that they don’t care about cars but love the hosts.}

*** ^{It was an interesting argument, but I’m not sure I buy it.}

^{Written by a human.}