AI podcast voices successfully fooled many listeners in university study

A university study at Oslo Met investigated how well people can distinguish between human and artificial intelligence generated podcast voices. Researchers Syed Fawad Ali Shah and Nusrat Akter, along with professor Frode Eika Sandnes, let sixty students listen to various podcast clips created by humans or using tools like Google NotebookLM and ElevenLabs.

While participants generally told the difference, artificial voices successfully fooled many listeners. In one fully AI-generated podcast created with NotebookLM, 56 percent of participants mistook the synthetic voices for real humans. Listeners sometimes also misidentified human voices as artificial. Details such as rhythm, pacing, lack of spontaneity, and missing emotional engagement in sports clips often helped people spot artificial speech, though synthetic tools are becoming increasingly realistic and capable of including conversational filler sounds like umm or eh.

The findings highlight risks related to voice cloning and fraud, where convincing fake voices could be used to impersonate acquaintances and request money or passwords. Researchers emphasize the need for a critical attitude toward audio just like other AI-generated content. At the same time, the study showed that a synthetic voice does not need to fool listeners into sounding human to be useful, opening up practical applications such as converting educational curriculum into audio summaries for students.