More and more technologies using artificial intelligence for podcast production have been developing in recent years. There is something for everyone, ranging from the complete creation of a real podcast to a complete roadmap for hosting a podcast or even translating a podcast into several languages.

NotebookLM’s phenomenon

One of the most popular recent tools is NotebookLM, an AI tool from Google introduced last year. It became viral over the past few weeks for its Audio Overviews option, which allows the creation of audio that mimics the speaking cadence of podcasters. The examples of generative AI podcasts have multiplied online, and the results are most of the time quite surprising.

Raiza Martin, who leads the NotebookLM team inside of Google Labs, boasts of the enormous possibilities, while it is possible to put up to fifty source documents in each Notebook. “The magic of the tool is that people get to listen to something that they ordinarily would not be able to just find on YouTube or an existing podcast,” Martin told Wired.

Editor’s note: to sample how NotebookLM works, listen to an AI-generated podcast based on The Fix’s article on French-language media outlets from the Balkans and Central Europe.

Laura Sibony, a French writer and teacher at HEC Paris who specialised in AI, recently tried it. She used it as an online research assistant to prepare her new podcast radio show. “We talk about the notebook for its functionality of managing podcasts from text, but it is already interesting for interacting with text bases. I needed to read a lot of different authors each week for my show, so I wouldn’t have been able to read everything. I relied on the notebook analysis, which does an analysis,” explained Sibony.

AI in podcasting

Notebook offers this option, just like many other applications that have developed to put AI at the service of podcasting. Large platforms like Spotify have, for example, already launched Voice Translation for podcasts. A groundbreaking feature powered by AI that easily translates podcasts into additional languages, all in the podcaster’s voice. Other applications, like Podgenai, allow users to produce episodes of about an hour on any subject in the blink of an eye.

The process is very simple. After choosing a theme, the AI analyses the topic in depth, identifies the relevant sub-themes, and generates a complete and structured script for the program. The application also has a mission to popularise technical concepts while making them captivating for the audience. The tool therefore directly transforms text into audio thanks to a good quality speech synthesis that allows users to choose from several voices.

With the multiplication of these artificial technologies, should professional podcasters be worried by this AI trend? According to experts, not really. Even if AI podcast tools bring many advantages, like engaging summaries and ideas, the synthetic voices will never be like humans.

Studies in different fields show a similar pattern: people may value AI-generated content, sometimes even more than human-generated content, but only if they don’t know the source
Pablo Sanguinetti, a researcher on AI and the humanities and adjunct professor at the Spanish IE University

“As soon as they find out that it is a machine and not a person, they tend to prefer the latter.”

Moreover, according to him, AI is not yet capable of reproducing unexpected human behaviors. “In the case of automated podcast production, I would lower, or rather redirect, initial expectations. Not so much because of the limitations of the model, but because of the nature of human behaviour, which is often more unpredictable and complex than companies would like,” added Sanguinetti.

Risks and questions

Artificial intelligence continues to raise questions about its use and its introduction in the podcast world. A big one is about the use of data and intellectual property. With NotebookLM, is it possible to load many texts even if we don’t have the rights to use it. “I think we need to rethink intellectual property law and generate new things whose status is not clear. Does generating text from an author database mean owing something to the copyright holders of the author?” wondered Laura Sibony.

Added to this are also ethical questions, links to bias. “An AI in itself is not racist, but this database may not be representative of an entire population. If the majority of the podcasts are made by men and young people, for example, it is a bias that will necessarily be reproduced with AI trained on already existing podcasts,” explained Sibony. Furthermore, as technology is not perfect, errors can occur, as can false information from unreliable sites.

We have also recently seen an explosion of podcasts on a variety of subjects, developed by the media but also by influencers or creators. Faced with an increasingly broad offering, which could accelerate even more with the use of AI in podcasts, we can wonder if the public will follow?

“We are going towards an erosion of podcast interest because there are too many. We will have a hyper personalisation that we lock ourselves in a bubble of filters. We will no longer look for things that lead us to think, to be in disagreement in refining one’s thinking through contradiction. The risk is to only go to podcasts that resemble us,” said Sibony.

Moreover, the future is always uncertain with this type of new technology, which will probably depend on the legislative framework that will surround it, and on the use we make of it. “AI-generated podcasts may now amaze us as a technical achievement. But they will only be commercially and aesthetically appealing if they serve to enhance human capabilities, not imitate or replace them,” said Sanguinetti.

Source of the cover photo: Ben Koorengevel via Unsplash

NotebookLM

Latest in NotebookLM

AI in podcasting: is it the future?