There are millions of podcasts in dozens of languages, so much so that you might think there’s at least one podcast episode on almost any topic. Of course, this is not the case given the breadth of human knowledge, but there is a system that has the ability to fill any gap in the verbal narrative of anything. It is called NotebookLM, it is inevitably based on artificial intelligence systems, and anyone can use it and produce a podcast of any type in just a few minutes, based on the documents provided to them.
It is far from perfect and sometimes leads to results that fall short of expectations, and it is no coincidence that the people who invented it at Google called it “experimental”, yet it gives an idea of how quickly systems of this type develop with all the implications that this carries. From this case.
LM notebook It has been around since 2023, but in its initial version it did not include the option to create a podcast. Its name is derived from the union of the words “notebook”, i.e. “notebook” in English, and LM which alternatively means “Language Model”, i.e. a model that works on a probabilistic basis for analyzing and producing contents, mostly textual. There has been a lot of talk about LM in recent years, especially after the progress achieved by the American company OpenAI with ChatGPT and then that acquired by Google with its own artificial intelligence system called Gemini, which began to be included in various online services provided by the company.
Systems like ChatGPT are known for answering any question they are asked, thanks to the vast amount of data the AI systems have used for training and the resources they can consult online. Specifically, because of the sources they use, and the way they put them together, the answers are not always very reliable or are completely wrong due to serious misunderstandings (technically called AI “hallucinations”). The ability to narrow down knowledge of these systems by providing them with a shortlist of reliable sources to work from can reduce the problem, and is an option often used, for example, in institutional or academic settings.
With this possibility, Google developed NotebookLM, proposing it as a system for taking notes more easily while doing research on a particular topic. You can upload sources in multiple formats, from PDF files to audio files, including YouTube videos and pages typically accessible on the web. The system analyzes all the information and automatically prepares a text summary, table of contents and a series of frequently asked questions study guide.
In addition to predefined answers, you can directly query NotebookLM AI through chat, thus asking more specific questions about a specific topic. Each answer is accompanied by notes and references to the parts of the sources from which the information was taken, so you can check and compare them. In some cases, the system also provides some additional contextual detail by indicating “its own” knowledge, noting that this is additional information that should be independently verified. Google says that the system does not use the documents provided to it for training, and therefore the data is used for the sole purpose of providing answers to the individual user.
NotebookLM is a simpler-to-use version of similar tools, allowing you to correlate large amounts of data and extract the required information. In the academic field, they are used to collect studies dealing with a specific topic, in order to compare methodologies and applied conclusions with the aim of producing new studies. In a corporate context, they can be useful for improving processes or sharing information and practices more easily between different departments.
After its launch in the United States in 2023, Google has been working on various updates to bring NotebookLM to other countries, including Italy, and enrich it with new features. About a month ago, the company added the “Audio Summary” feature, which allows you to produce an “in-depth conversation” about the topic you are looking for. The audio file is created within a few minutes and is structured as a conversation between two podcast hosts, who explain and discuss the contents of the sources provided to the system.
The conversation is essentially designed to be a short episode of a podcast, with a male and female voice speaking, interrupting each other, sometimes speaking uncertainly, with pauses and other imperfections that make the end result more realistic. Each episode follows roughly the same schedule: there’s an introductory segment in which the two hosts frame the topic, often with examples and wasting little time, a central segment in which they present the main topics and finally a conclusion that summarizes the meaning of the conversation. The quality of conversation varies greatly, and also depends on the type and quantity of resources uploaded to NotebookLM, and the only language currently available is English (but resources can be used in almost any language).
The audio below is the beginning of a conversation generated by the AI system based on the first version of He explained things wellBook Magazine mail, Dedicated to gender issues. The feeling is to hear from someone (something?) who has internalized the book’s meaning, goals and approach to very sensitive and sometimes polarizing topics.
The system allows you to obtain texts even from sources that do not have strictly discursive or informational contents. In the case below, the conversation is about a recent version of Official Gazette of the Italian Republicthrough which the regulations in force in Italy are published and announced. The conversation begins by announcing that we are going to talk about “Italian bureaucracy,” adding immediately afterward, “But before you click away and think ‘how boring,’ trust me, it will get interesting.”
In roughly ten minutes of podcast, not all of the topics in the podcast are normally covered Official Gazettebut the most relevant ones have been highlighted. This approach is considered too discursive and, moreover, the hosts often remain generic and there is no possibility to customize the vocal result according to one’s needs. For this reason, above all, the system did not arouse much enthusiasm among some researchers and IT experts, who tested it in some of their studies.
For example, Professor of Computer Science at Brown University (USA). I mentioned The system seems to behave like a “novice researcher: it generally understands the meaning of what is being talked about, but does not always know what to focus on.” In a test of scientific research, for example, “part or all of what makes that study relevant is missing.”
Therefore, the conversations produced by NotebookLM cannot be used to facilitate the study of a topic, compared to other tools provided by the system. However, at the same time it retains a certain charm because the system has the task of producing an audio file anyway, which shows that it is able to emulate a certain creativity. The audio below was produced, for example, by sending the AI system a PDF of the menu of a fast food restaurant that sells kebabs in Italy. Although the source only contains the names of the dishes and their prices, the two voices manage to carry on the conversation, although it sounds off-putting at different times.
It is likely that in the coming months the conversation production system will be expanded to not only improve its performance, but also to add other languages such as Italian and reduce the amount of strangeness caused by translation from our language to English. Perhaps the best way to get an idea of the hosts’ reliability is to listen to their conversation below, which was produced using the article you just finished reading as the source.
“Incurable internet trailblazer. Troublemaker. Explorer. Professional pop culture nerd.”
