I find Google’s NotebookLM to be one of the most powerful tools for researchers these days. Its core function is to work with the material you upload. Hence, it differs from ChatGPT and other tools that invent things on their own or search the web. It also has many different types of reports, including audio and video. But how accurate is it?
Testing NotebookLM
To test NotebookLM on some material I know very well, I decided to provide it with a collection of public data about myself, more specifically, what I have registered in the Norwegian NVA system, which is the database we use in Norway to register all our academic activities. This includes academic publications, public lectures, artistic works, interviews, etc. In my case, that contains around 1000 entries dating back to 2000.
I began by exporting all my entries from NVA as a .csv file. NotebookLM doesn’t read CSVs, so I quickly converted the document to a PDF. That file looks like a mess to me, but since it’s based on structured data, I assumed NotebookLM would figure it out. And, sure, it did!
Writing a bio
Academics have to write bios about themselves all the time. I have one on my personal UiO page that I adjust from time to time. In the past, it was always a pain to adjust the bio to various requirements (typically 50, 100, 200 words). Nowadays, I find LLMs helpful in changing the length of such bios; still, that requires you to have a bio to start with.
It can be fun to ask ChatGPT or CoPilot to write a bio about yourself, but in my experience, it doesn’t work very well. Even though they come up with many correct things, they often fail to provide a good narrative. And, at times, they are incorrect. But how, then, will a bio based on my NVA report look like? After all, that is a complete list of all my academic “results” over the years. On the other hand, there is a lot of information about me that is not covered in that material. Still, given that quite a limited starting point, NotebookLM manages to rewrite a remarkably accurate 200-word bio:
Alexander Refsum Jensenius is a Professor and the Deputy Director at the RITMO Centre for Interdisciplinary Studies in Rhythm, Time and Motion at the University of Oslo. He earned his Master’s degree in 2002 followed by a PhD in 2008, where his research on air instruments earned him the nickname “Dr. Luftgitar”. His academic work focuses on the complex relationships between musical sound and human body movement. He is particularly known for his studies on music-related “micromotion,” including the “National Championship in Standing Still,” which explores involuntary responses to sound. Jensenius developed the Musical Gestures Toolbox and is a central member of the MishMash Centre for AI & Creativity. Beyond musicology, his movement-analysis techniques have been applied clinically to identify ADHD and cerebral palsy in infants. He is a prominent advocate for Open Science and has been instrumental in the development of the NOR-CAM framework for research assessment in Norway. As a research musician, he performs using interactive systems, electronics, and novel interfaces like musical balls and joysticks.
That is actually very good. It has made a few factual errors, but those are understandable. For example, I was the deputy director of RITMO but took over as director in 2022. Those titles don’t figure often in my publications, though, so it is an understandable mistake. I see that it also calls me a “central member” of MishMash, but hasn’t figured out that I am the director. That is also understandable, given that NotebookLM only has a bunch of titles from various lectures I held on MishMash last semester, but doesn’t know anything about the initiative otherwise.
In summarizing my research contributions, it does a remarkable job. It accurately summarizes what I would consider to be my core contributions. It is also interesting that it picks up on Open Science and NOR-CAM, which is understandable given that I have given quite a lot of public lectures on those topics.
All in all, this is a bio I could actually use with only minor modifications!
MindMaps
I find mindmaps fascinating, and I have even taught using mindmaps. This is what NotebookLM comes up with based on my NVA report:
It is brief but captures the results well. I like that it includes “Education and Outreach” as a separate category.
Infographic about myself
The ability to create infographics is one of my preferred features of NotebookLM. Now, I was curious to see how it visually summarizes my academic career:
After a few decades in academia, I have been doing quite a lot of different things, and it is not always easy to know how to present myself and my activities. I think it nicely picks up some of the core topics, with nice illustrations of micromotion, motion capture, AI, and Open Science. The air guitar is visualized as a real guitar, though, but that is the only real “problem” in this visualization.
Presentation
Equally fascinating as the infographic is the multi-slide presentation it makes for me:
I think these slides build a nice story that covers many of the aspects I have been concerned about: “Music and Movement”, “Technology and Tools”, “Institutions and Leadership”, and “Open Science and Systemic Change”. In my own slides presenting myself, I typically include four points: “music researcher”, “research musician”, “centre/lab director”, and “Open Science advocate”. NotebookLM takes a slightly different approach, but it is a good summary of what I have been working on.
Video
NotebookLM has also made a narrated video for me, called the Sound’s secret influence:
Interestingly, it picks up on two topics: “Decoding the ASMR tingle” and “The mystery of the beat”. Why did it choose those? First, regarding ASMR, I only have one publication on this topic, which was based on Henrik Sveen’s Master’s thesis that I supervised last year. So it is a marginal contribution of mine compared to many other things. Also, I would argue that ASMR is a niche (but interesting!) research topic; I don’t think many musicologists or psychologists have even heard the term. So why did NotebookLM pick out this particular contribution from my 1000-entry-long list? My best guess could be that it has gained some popularity on YouTube in recent years. Given that my other research interests are also relatively niche (micromotion, standstill, ventilation sounds, etc.), is ASMR the topic with the broadest public appeal after all?
In any case, NotebookLM didn’t have much information to go from when making the video. All that is provided in the NVA report is the abstract of the paper:
Autonomous Sensory Meridian Response (ASMR) is a tingling sensation in the neck and spine often triggered by specific sounds. This paper reports a study on the impact of different cyclic patterns and spatial orientations—defined here as the perceived directionality and motion of sound sources in a three-dimensional auditory space—on inducing ASMR experiences. The results demonstrate that both the type of cyclic pattern and the spatial orientation significantly influence the intensity and nature of ASMR experiences. Furthermore, the research explores synthesizing ASMR-inducing sounds while preserving key audio characteristics from acoustically recorded ASMR content. Through survey data analysis and regression modeling, distinct patterns emerge regarding the relationship between personality traits and ASMR experience. The findings contribute to a deeper understanding of ASMR as a sensory phenomenon and provide insights into the potential applications of artificially generated ASMR stimuli. Additionally, the research sheds light on the role of spatiality in ASMR experiences and the synthesis of ASMR-inducing sounds for future studies and practical applications
Given this limited input, it manages to squeeze in a couple of relevant slides with narrated voice into the presentation.
The second part of the video is based on some of my standstill work. Here, I have many entries in the NVA report on this topic, and it appears to combine multiple findings into a single point. The summary is quite good, although somewhat exaggerated. For example, it claims that the impact of music on a standstill is much larger than we have shown.
From this video, I can see that it also “hallucinates” to a certain extent. There is no mention of “rock vs jazz” in the NVA report, so this is something it has invented on its own. It is not a bad metaphor for “simple” versus “complex” rhythm, though, but it is always worth being careful about LLMs inventing things that is not in the original content. I will keep an eye out for this in future testing.
Podcast
The ability to make podcasts was one of NotebookLM’s distinguishing features when I began using it last year. These are in the form of a “dialogue” between a “male” and a “female” voice:
It starts by focusing on micromotion, but then moves on to include numerous other examples from my research outputs. One reason it works so well is that it only has titles and abstracts to work from. Those are already distilled versions of my research, written to be relatively easy to understand.
Since some of my work is in Norwegian, it confuses some titles and adds strange pronunciations of “Musikk er bevegelse”, “doktor luftgitar”, and “MP3-spiller”. I also find the tone very glamorous, using terms like “revolutionary” and “amazing” that I think are overused.
In general, though, I am amazed (to add some “glam” myself) about how it manages to create relevant references, for example, mentioning “motiongrams” in the discussion of data and methods sharing in NOR-CAM.
I also like how it explains motiongrams with a “heatmap” metaphor and describes motion capture by referring to “Hollywood animations”. Here, it clearly hallucinates again, because this is not something in the provided material.
Given all our internal discussions about the MishMash name, I find it interesting that NotebookLM likes it, which points to our ambition to bridge perspectives.
NotebookLM’s liking for ASMR leads to ending the podcast on this topic and a forward-looking excitement about “microengineering soundscapes”. That is not a term I have used before, but I like it!
In sum
Over the past year, I primarily used NotebookLM to summarize various academic texts. It is remarkably good at capturing the essence of large collections. It hadn’t occurred to me that it would also be able to parse a “dataset” like my NVA report. But why not, after all, it is also a collection of textual elements that can be analysed. And, as the examples above show, it accurately captures key parts of my academic life over the last decades.
The new “multimodal” features are also interesting, including the ability to create infographics, presentations, videos, and podcasts. Interestingly, they all pick up on different things; it is not the same story being told in a different format. As far as I can see, you will also get new versions each time you try, which makes it possible to co-create with the system as you go.
All AI-based systems have biases, and so does NotebookLM. It shows particular interest in ASMR, for example, and adds US-based explanations (like the Hollywood analogy). In general, though, it does a good job of summarizing the material. The mistakes it makes are understandable, given the content on which it is based.
A cool feature is that you can share the NotebookLM notebooks; check it out here if you are interested. Since it is changing frequently and might even disappear in the future, I have provided the downloaded files above to preserve them as a historical record.
I, of course, used NotebookLM to generate much of the content presented here, although all the writing is my own, corrected by Grammarly.