AI voices are getting too real to spot, study warns
A study published in PLOS One shows that speech synthesis has advanced to the point where listeners often can’t tell a real voice from one generated by artificial intelligence. The effect is strongest with voice clones, AI-made copies of real people that sound almost identical to the originals, Kazinform News Agency correspondent reports.
Researchers tested how people perceive three types of voices: natural human speech, fully AI-generated speech, and “clones” based on recordings of actual speakers. Volunteers were asked to rate how real the voices sounded and to decide whether they came from a person or a machine. They also judged how trustworthy and dominant each voice seemed.
In many cases, the clones were so convincing that they were taken as “more human” than the original recordings. On average, 58% of the cloned samples were mistaken for genuine human voices. Even fully synthetic AI voices fooled listeners about 40% of the time.
People often mistook genuine human recordings for artificial ones. In some rounds of testing, only 62% of real voices were correctly identified as human. In other words, the sense of authenticity shifted depending on context, next to synthetic samples, real voices started to sound “less real.”
Consequences
The study also revealed a social effect. Listeners rated AI voices as more dominant than human ones. In some cases, they even seemed more trustworthy than actual speech. This suggests that artificial voices can do more than blend in. They can shape impressions and influence how people respond.
AI-generated voices are increasingly common in advertising, podcasts, and voiceover services. But once they become indistinguishable from real ones, the risks of fraud and manipulation grow. The researchers stress that people tend to attribute human qualities to machines, which could lead to misplaced trust in artificial voices.
Earlier, Kazinform News Agency reported that study revealed who turns to Gen-AI most in higher education.