Notebook LM

 .

These podcasters aren’t real: Our AI experiment with Google’s Notebook LM

I think my colleague Björn Hedin told me about a new AI tool from Google, Notebook LM, sometime in (possibly the second half of) October. Someone had notified him about the tool and Björn had then created an audio file that summarised a phd thesis into a 30-minute long "podcast". We both tried Notebook LM some and thought it was extremely interesting; you can upload a pdf file of a scientific article to Notebook LM and it will generate a "podcast" (sound file/audio overview) where two podcasters who know each other well together walk you through the main results of the article and also, well, banter in such a way that the walk-through can become quite entertaining (but for some instead "too chatty"). And it only takes a few minutes for Notebook LM to generate such an podcast out of thin air. There are of course no real podcasters, rather "just" an AI that "has listened to" (been trained on data from) a zillion hours of TV, radio and podcast hosts who talk to each other. This very powerful tool can however basically create a "podcast" out of any file so that you can listen to an audio summary that highlights the main points of an article, a bunch of articles, a powerpoint presentation or a whole phd thesis:

"The idea behind the “Audio Overviews” feature is simple: take a bunch of documents, websites, YouTube videos, etc, and generate a podcast out of them."

Not long thereafter (on November 1), a student sent me a mail and asked me to make my lecture slides available before the lecture or at least directly after the lecture so that he could use Notebook LM to create a podcast that he could listen to on his way to the lecture or at least on his way back home. It just so happened that Björn would give a guest lecture in my course just 10 days later and part of the literature for that lecture was a systematic literature review that Björn, me and our colleagues Cecilia Katzeff and Elina Eriksson have written:

Hedin, B., Katzeff, C., Eriksson, E., Pargman, D. (2019). "A Systematic Review of Digital Behaviour Change Interventions for More Sustainable Food Consumption". Sustainability 2019, Vol. 11, Page 2638, 11(9), 2638 (available online)

As is the case for all systematic literature reviews, this review is quite boring but very useful, so we decided to do an experiment. We created a 13 minutes long Notebook LM podcast of the paper (available here) and made it available together with the article itself. We unfortunately didn't think about following it up by for example asking the 135 students who took the course if they had listened to the podcast and/or read the article and what they thought about them. 

What did happen though was that I immediately followed up the lecture with a message to all the students in the course where I encouraged them to get in touch with me if they were interested in writing a Notebook LM-related thesis during the spring term. There was a lot of interest and now, 10 weeks later, it is clear that I will lead a supervision group where eight students will work in pairs (at a 50% pace) throughout the spring term and write their bachelor's theses about Notebook LM. I'm really excited about this and think the results could be very interesting. I will also continue to do my own experiments with Notebook LM (as will Björn) independently of what "my" thesis students will do during the spring. 

One of the issues we are interested in is the accuracy of the podcasts that Notebook LM generates and we will therefore reach out to researchers in different disciplines with an offer, so if you happen to be a researcher and you are interested in being served with podcasts of scientific articles (we are primarily thinking about podcasts that are based on your own texts!), then by all means feel free to reach out to me by mail.

It might sound as if I'm super positive about Notebook LM, but that is however not  the case. I'm quite worried about vario us possible or plausible negative effects of using the tool. While it's extremely powerful, it's not clear what the effects of using it will be, and that might then be part of what we want to explore during the spring. My uneasiness with the tool is not shared by my colleague Björn who has more charitable view of human beings (including of students) than I have. I worry about whether students will stop reading articles – since it's much easier to listen to a 10-15 minutes long podcast than to struggle with a text that it might one or two hours to read. When then might be the consequences of that be? Well, see for example my text about Nicholas Carr's book "The Shallows: What the Internet is doing to our brains" that I wrote about (on this blog) 14 years ago:

"When deep reading gives way to shallow reading - involving skimming texts, following links and skitting from text to text and channel to channel - Carr also worries about deep thinking giving way to shallow thinking. [...] In Carr's words we become "pancake people" who know little about much."

Perhaps I worry too much, but then again perhaps I don't worry enough...? The only thing I'm sure of at the moment is that I will have a better grounding for being worried/not being worried four or five months from now... and I will certainly get back to this subject later this spring when these investigations are finished!

Here a few links to resources about Notebook LM (from Wikipedia):

• Latent Space (65 minutes long podcast), "How NotebookLM Was Made" (October 25, 2024)
• Levy, Steven, "Google’s NotebookLM Aims to Be the Ultimate Writing Assistant" (Wired, Dec 8, 2023)
• Wang, Biao (product manager at Google Labs), "NotebookLM now lets you listen to a conversation about your sources" (Sept 11, 2024)
• Boran, Marie, "Five Productivity Hacks Using Google's New AI Podcast Creator" (Newsweek, Sept 23, 2024)
• Orland, Kyle, "Fake AI “podcasters” are reviewing my book and it’s freaking me out" (Ars Technica, Sept 23, 2024)

.