NotebookLM
What are it's use cases?
Google NotebookLM has been around for a little while (at least by the high-paced standards of AI product development), but has recently moved from ‘Beta’ to ‘Experimental’ and also been made available outside the USA for the first time.
My Linkedin feed has been filled with people discussing this, and having now toyed with it myself, I have been piecing together my thoughts on what I think the best use cases for this product are. This will be discussed in light of academic literature about various media that can be used in a university course. My conclusion, is that this has some great uses, though some of the uses suggested so far are more likely to be gimmicks, at least until new functionality is added to the product.
What is it?
Helpfully NotebookLM has a suggested prompt on the first page called ‘Introduction to NotebookLM’ that guides you around the product whilst using it. This introductory notebook has several sources,. including a ‘getting started’ document and a glossary. It also includes some notes, and suggested questions to get the conversation with the AI started. The key features of this product are:
It is powered by Google Gemini 1.5 Pro, the companies leading model, and peer-competitor with the other leading models from OpenAI, Anthropic etc.
Each notebook can hold up to 50 sources. These can include PDF files, Google sheets and docs, links to websites (where the text on a specific page is read), and YouTube videos. You can also copy and paste chunks of text from other sources.
All of these sources can then be synthesised and examined as a collective. As an example I copied the links to all of my blogposts to date into an ‘Academy Empirica’ notebook. With this I can then ask questions of all the material from my blog. You don’t have to use all sources however, you can select a subset if you you have a reason to focus on something specific, using simple tick-box controls on the lefthand panel of the notebook.
The Notebook, like many other AI products, has suggested questions to get you going. For my Academy Empirica notebook I see suggestions like “What are the potential challenges and limitations of using AI in the classroom, and how does the author address them?”
The output from such queries includes clear references to where that point is sourced from, so you can track and check where things come from (as it notes at the bottom of the page, this is an experimental AI product and can make mistakes.)
This output can be saved as ‘notes’ in the way that LLMs save your chats, but these are attached to the notebook for further reference.
By clicking on ‘Notebook Guide’ in the bottom-right, I get a summary of all the material in the notebook, and suggestions on documents I can create, which include study guides, FAQs and briefing documents.
Podcasts. Perhaps the most eye-catching element of this software is the ability to turn your notebook into a podcast. This is one of the suggested creations in the ‘Notebook Guide’ area, and the outcome is pretty amazing. Here is the podcast version of all my blog posts to date:
What is the best use for Notebook LM?
To understand the use of this product, it is best to understand the difference to other AI products.
Firstly there is the exploitation of Google Gemini’s larger context window. This is the amount of information that can be processed by the AI at once, and is an area that Google have been ahead in compared to other AI chatbots. This Notebook format combines that large amount of background information with an easy to manage style whereby you can create Notebooks on specific topics, and use subsets of information in an easy way. So I can load every lecture and every workshop from one of my courses, and either use all of it, or just the lectures, or just the workshops, or just the lectures and workshops on a specific topic, all in a pretty user friendly way.
Specific uses within an educational setting include:
Creating study guides, handbooks and other module materials. I tried uploading all my lectures and workshops for one of my modules, and hit the button for ‘study guide’. The exact style was not what I wanted, but it generated a quiz, suggested answers to the quiz, some essay questions and a glossary of key terms, all of which seemed accurate. The possibility to speed up the production of documents is clear, and the advantage of this specific product is the amount of material that can be used and the interface for using it.
For the same module, I used NotebookLM to generate ideas for exam questions, this time with my own prompt including more detailed style instructions, and also to generate model answers. The quality was good, and this has the advantage of (almost) guaranteeing that your exam questions are based on the material of the course. Anyone experiencing student complaints (or being the student complaining) about exams that don’t match what was actually taught will be aware of this importance.
Remi Kalir has some great ideas about how to use Notebook LM. These include using the notebook as a guide for instructors who are new to teaching a course, for using to integrate student feedback into course design, and for sharing with students as a way for them to work with the course material. This last idea is a version of what I have tried to do previously with my course GPT, but the ability to include more material makes NotebookLM a strong contender for this task1.
Similarly Eric Hudson has good ideas on using AI, including NotebookLM as a writing assistant. His conclusion is that AI is a “Powerful assistant and a Poor Replacement”, which seems fair. One of the assistant tasks NotebookLM appears helpful with is getting constructive criticism, helped by its abilities to reference where it’s points are coming from in a way that other AI chatbots do not.
Podcasts: This was the feature of NotebookLM that appears most unique, and it has a mixed response. Andrej Karpathy is a fan, highlighting that the user isn’t obligated to think of a question to get started understanding something. Zvi Mowshowitz is a little less convinced, but does note that listening to his own creations allowed him to spot errors he had made, and get a better sense of some of his own arguments. Sean Thomas predicts that this could be a killer of actual podcasts and the people that make them. This feels a long way off, but with the pace of advance, doesn’t seem very unlikely in the medium-term.
For myself, though the output is amazing, I didn’t feel like I wanted to listen for too long. This may just be a personal preference for reading over listening, which will be different for other people. I also dislike the tone of the podcast in terms of voice and mannerisms.
Analysing some more formal academic work on the topic, we see more reasons to be doubtful of the utility. Philippa Hardman notes that using AI to summarise in this manner, whilst useful as an introduction to a topic, also misses much of the complexity. She also emphasises that the process of learning is complicated, and we potentially lose that if we engage only with these shallow summaries:
More formal empirical evidence shows reason to doubt the efficacy of podcasts as a learning tool also. This study by Choi et al (2024) finding no significant improvement from use of podcasts.
Weaknesses
To add to these doubts, there are some other marked weaknesses in this product.
Being a Google product, NotebookLM is easy to integrate with Google documents of various types, but more difficult if not. I had to convert all my PowerPoint slides to Pdfs to use it, which took some time. It also does not output documents in a very user friendly way, needing a little editing in the case of my study guide experiment for example.
As mentioned already, the podcasts are all in the same style, with no choices. I look forward to options to change voices and the style of the podcast (yes I am being demanding of this miraculous technology to which I contributed nothing!)
As with all AI products in this space, there are issues around privacy, copywrite etc. I find myself being too relaxed about what information goes where, and how I use that, which can lead to troubling outcomes.
There is a lot of new evidence coming out about AI-use in education, though the timelines of academic publication do not fit well to the fast-paced world of AI-development. It will be good to see more evidence comparing student cohorts using these tools with those that do not, to get a better sense of the strengths of this approach.
Conclusion
NotebookLM is a powerful tool. For now it is free to use with a Google account, with no restrictions on the power of the model or features. As with other AI-methods of summarising or interrogating a large number of sources, the notebooks you can produce have an ability to know your work better than you yourself know it, and do similar for the work of others, though in a way that may miss some nuance and complexity. The quality of outputs is not perfect, but is generally excellent, and could have several beneficial uses in the workflow of an academic.
Whilst the podcast production ability was perhaps the most eye-catching part of this product, it is not one that I will to use too much for now. Once these become more customisable then I will see greater uses for them, and many people will like the option for an audio version rather than relying on written summations.
As with any AI product, the best way to understand what it can be useful for is to use it yourself, enter some of your course materials or a selection of your written material and explore what you can do with this product.
Google’s NotebookLM joins ChatGPT, Claude, Perplexity and others as a useful AI-tool that can help students and academics. None of these is a one-stop shop for everything you want, and blending the different models, using each where it is strongest, will be the path to greatest gains for now.
A weakness of this approach appears to be that sharing a Notebook works mainly through personal Google accounts, so this may be more difficult with students where we may expect to operate through official university accounts. There is also a limit of sharing for 50 people, so for now this use-case is limited.