A groundbreaking study has utilized artificial intelligence (AI) to create podcasts that summarize research from scientific articles. Remarkably, many authors believed that the voices behind these podcasts were human.

This research, shared in the European Journal of Cardiovascular Nursing (EJCN), was led by Professor Philip Moons from the University of Leuven in Belgium. The team employed Google NotebookLM, an AI tool developed by Google, to craft podcasts explaining recent research featured in the EJCN.

During his presentation at the Association of Cardiovascular Nursing and Allied Professions (ACNAP) conference in Sophia Antipolis, France, Prof. Moons remarked, “In September 2024, Google introduced a new feature in NotebookLM for generating AI-based podcasts. This got me thinking about its potential benefits for researchers and editors.”

“My initial test with one of my papers left me amazed by the high quality and natural sound of the output. I realized then how valuable this technology could be for scientific communication. However, we needed to assess its quality and accuracy, which led to this study.”

Prof. Moons, who is not only a health science educator at KU Leuven but also the editor-in-chief of the EJCN, selected 10 diverse articles and sought permission from their authors to participate in the study without revealing that the podcasts were AI-generated.

The team sent the AI-created podcasts back to the authors for feedback on engagement, credibility, and detection of AI involvement. Following this, they conducted a questionnaire (crafted with the help of ChatGPT) and a 30-minute interview over Microsoft Teams.

Authors praised how well the podcasts articulated their research in “simple and understandable language,” highlighted key points effectively, and maintained a good structure. Many believed the AI “hosts” sounded professional, with some anticipating that they had backgrounds in healthcare. The engaging manner of the speakers also stood out as a positive feature.

Most authors found the podcasts to be trustworthy sources of information but pointed out issues like the American accent, certain phrases that seemed overly enthusiastic (words like “amazing” and “groundbreaking”), inaccuracies, and incorrect use of medical terminology. They emphasized the importance of thorough fact-checking before release.

AI detection

When asked, “Would you be surprised if I told you this podcast was generated through Artificial Intelligence?”, five authors responded with surprise, including one who had previously studied AI. Their reactions varied from “shocked” to “having an existential crisis.”

Five authors were not surprised. They suggested that if AI-generated podcasts were released, it should be clearly stated, with due credit to the authors and a reference to the original paper.

The participants generally agreed that these podcasts would be most beneficial for patients and the public, given their straightforward delivery. Some also noted that they could help readers keep up with the latest research and increase the visibility of original articles.

As the technology evolves, some authors suggested customizing podcasts for specific audiences based on demographics or interests. Currently, the AI cannot change voice styles or languages, but these features are anticipated in the future.

Prof. Moons remarked, “The overall accuracy of the podcasts is impressive. Since we are just scratching the surface of AI-generated content, we expect the quality to improve vastly in the coming months. These podcasts seem particularly useful for non-specialists, like the general public or patients.”

“AI-generated podcasts could revolutionize communication in science. They allow for easy production by simply uploading articles and possibly providing a few prompts. This could become a sustainable method to present information to those who typically don’t engage with scientific journals.”

He added, “Even with this technology, human podcasters will still have a vital role. There will always be topics requiring a human touch for accurate and sensitive treatment, and I foresee a blend of AI and human effort in future podcast episodes.”

Looking ahead, the researchers plan to further explore the potential applications of these AI-generated podcasts, including gathering public opinions and examining their use during scientific conferences.

“For instance, we could create podcasts summarizing the content from ESC conference sessions for those unable to attend,” Prof. Moons suggested. The researchers have also produced an AI evaluation of their findings, accessible through a link in their published paper.