AI-Generated Podcasts Bridge the Gap in Science Communication

AI-Generated Podcasts for Science Communication

Scientists have long understood the need to communicate their research clearly to the public. However, many researchers lack formal training in communication, making this a challenge. Podcasts have become a popular and effective format for sharing scientific findings—they offer on-demand access and reach diverse audiences.

To explore new ways of using this medium, a team in Belgium tested whether artificial intelligence (AI) could generate podcasts based on scientific papers (Eur. J. Cardiovasc. Nurs., doi: 10.1093/eurjcn/zvaf074).

Creating Podcasts with AI

Philip Moons and colleagues from the University of Leuven used Google NotebookLM, an AI research assistant, to produce 10 podcast episodes. They selected this tool because it links generated content to specific sources, which helps reduce errors or hallucinations—common pitfalls where AI fabricates incorrect information.

The podcasts covered a variety of article types from the European Journal of Cardiovascular Nursing, including original research, reviews, patient perspectives, methods, and discussion papers. The topics themselves were not a factor in selection.

Using the Audio Overview feature launched in September 2024, the team instructed the AI to produce “deep-dive conversations.” This feature creates a conversational summary tailored to an audience or focus area based on prompts. The podcasts ranged from five to 17 minutes. While eight episodes were accepted as generated, two required redoing with clearer instructions—one with a time limit, the other focused specifically on cardiovascular nurses.

Evaluating Quality and Trustworthiness

After generating the podcasts, the researchers sent each episode to the author of the corresponding paper without revealing its AI origin. The goal was to assess three key factors: engagement, trustworthiness, and whether the AI nature could be detected.

Authors found the podcasts engaging, appreciating the simple language and conversational tone. Some listeners even mistook the hosts for professionals with medical or nursing backgrounds. Most were satisfied with the pacing and length, though one mentioned that about 20% of the content felt like filler.

Trustworthiness scored well overall. However, some inaccuracies appeared, such as confusing heart failure management with diagnosis and mispronouncing scientific terms. Occasional overstatements like “amazing” or “groundbreaking” raised concerns about hype. The American accents and lack of host introductions also made some listeners question the credibility.

Only half of the authors were surprised that the podcasts were AI-generated. Those who suspected AI presence noted the absence of natural speech fillers like “um” and occasional incoherent moments as giveaways.

Potential and Next Steps

The study suggests AI-generated podcasts can be valuable tools for spreading scientific information to broader audiences with minimal effort. However, accuracy checks by the original authors are essential before release. Transparency about the AI origin and clear references to the original research should also be standard.

As the technology improves, AI-generated podcasts could offer an efficient way to communicate complex research to people who may not regularly engage with scientific journals.

For professionals interested in integrating AI tools into science communication, training resources are available through platforms like Complete AI Training.