Sophia Antipolis – For the first time, a study evaluated the use of artificial intelligence (AI) to generate podcasts, via the Google NoteBooklm application, from scientific articles published in theEuropean Journal of Cardiovascular Nursing (EJCN). The results seem convincing, since only half of the authors of the articles in question have perceived that the podcasters were not human.
This study aimed to determine the potential of this tool capable of generating a podcast simulating with great realism an exchange between two interlocutors on scientific work. The results of the study were presented at the congress of theAssociation of Cardiovascular Nursing and Allied Professions (ACNAP 2025). They were published simultaneously in theEjcn [1,2].
Too fluid communication?
The podcasts had an average duration of 10 minutes. Invited to assess them without being informed of their design mode, the authors estimated that the key points of their articles were summarized in simple and accessible terms, and in a structured manner. Some have even indicated that speakers seemed to have skills in nursing or medicine.
Participants noted, however, that the animators of the podcast spoke too fluidly, without imperfections, that is to say without hesitation, repetition, nor spontaneous interruption or reformulation, characteristics of a normal human conversation, which seemed unnatural. It is this smoothed aspect that led some to suspect the involvement of AI.
Half of the authors asked have also noted errors in pronouncing medical terms and inaccuracies. For example, a podcast has focused on the diagnosis of heart failure, while the article focused on the management of heart failure. Another podcast also talked about women solely, while the study also included men.
The researchers were also arrested by the American style of the speakers, who abused superlatives to describe results, which did not deserve such a craze, stressing in passing the need to adopt a more academic style. Note that the system now offers other languages, including French, also very credible.
Potentially useful for researchers
In total, the Pr Philip Moons (University of Louvain, Belgium) and his colleagues generated ten podcasts from articles from the journal European Journal of Cardiovascular Nursing devoted to care in cardiology. They concluded that “the podcasts generated by artificial intelligence are able to summarize the key results in a clear and accessible manner”.
Researchers believe that this tool already appears sufficiently advanced for relevant use between specialists. “Podcasts [générés par IA] are adapted to scientific communication, mainly for patients and the general public, but could also be useful for researchers and health professionals. ”
“What struck me is the general precision of podcasts. Knowing that we are only at the very beginning of this kind of technologies, the quality will most likely improve in the coming months, “said Professor Moons, in a press release. According to him, this tool could help researchers better share the results of their work.
Professor Moons explained that he had the idea of this evaluation after the launch by Google in September 2024 of the Notebooklm feature. “During my first test, carried out with one of my own articles, I was blown away by the quality and naturalness of the result. At that time, I achieved the potential of this type of tool for journals with a goal of scientific communication. »»
After having generated the ten podcasts lasting 5 to 17 minutes, the authors of the selected items were invited to assess the quality of the podcasts, by filling out a questionnaire and by giving their impressions through an interview of 30 minutes by videoconference.
Lack of context
Asked about the precision of podcasts, the participants all considered that the key elements of their research work had been taken up, in “a very simple and easy to understand form”. Another valued aspect: the presentation in the form of a conversation between two interlocutors which makes it possible to more easily capture attention.
Some welcomed the professionalism of speakers. “I was curious to know their journey, because it would have been said that they had received training in nursing or medicine,” said a participant. Many have been disrupted by the lack of presentation of the identity of speakers and contextual in the creation of the podcast.
Most have judged the contents of reliable podcasts. Nevertheless, in addition to the few errors of pronunciation noted, some have reported inaccuracies. For example, one of the authors interviewed pointed out that obesity was presented in a podcast as a “habit”, which “gives the impression that it is a choice”.
That said, half of the researchers solicited, including an IA expert, could not identify podcasts as being generated by AI. Some said they were “shocked”, “amazed”. Note that the majority of them are regular podcast listeners. Those who suspected the involvement of the AI also said they were also surprised by the result.
Increase the visibility of publications
All agree on the need to indicate that the audio was generated by IA. For the majority of the authors interviewed, the quality could be improved by adopting a more academic tone, especially if the podcast is intended for researchers. It has also been suggested to describe the methodology more and to highlight certain limits of the study.
If patients and the general public are presented as the “most appropriate target group”, these podcasts can be an “economic and effective” means for researchers and health professionals “to learn about the latest research advances”, increasing the visibility of publications, estimate the authors of the evaluation.
“This could become a lasting model to transmit messages to people who do not read scientific journals,” said Professor Moons, who nevertheless considers important to keep the participation of humans, in particular “to bring nuance”. According to him, the model could evolve towards a hybrid form, a podcast generated by AI, but in which the human could participate.
A perspective already put into practice by Notebooklm, which recently offers in its beta version (only in English) to interact with AI. Once the podcast has been launched, it is possible to exchange with one of the speakers by asking it. The answer is generated automatically and formulated orally, then the podcast continues after joining the human/machine interaction.
Follow Medscape in French on Bluesky , Facebook , Instagram , Linkedin , Youtube .
Register for newsletters DE MEDSCAPE: Select your choices