AI-cloned voices indistinguishable from real voices in study
A study published in PLoS One found that most people cannot distinguish between AI-cloned voices and real human voices, with 58 percent incorrectly identifying cloned voices as authentic. Researchers noted that commercial voice cloning tools are widely available and can be misused for fraud and impersonation without technical expertise.
What happened
Researchers at Queen Mary University of London conducted a study published in PLoS One where participants listened to 80 voice samples—half real, half AI-generated—and rated their authenticity and characteristics.
The study tested two types of AI voices: generic voices created from scratch and cloned voices generated from real human recordings. While 58 percent of cloned voices were incorrectly identified as real, 62 percent of authentic human voices were correctly recognized as human, showing minimal difference in detectability.
Reported impact
Dr. Nadine Lavan, the study's lead author, stated that AI-generated voices, particularly voice clones, sound as human as real recordings and can be created using commercially available tools without technical expertise or significant cost.
Voice cloning technology has raised concerns in entertainment, where well-known individuals' voices have been used without consent. Scarlett Johansson publicly accused OpenAI of using a voice that 'eerily reminded' her of her voice from the film 'Her'. Voice deepfakes have also been used to imitate politicians and journalists.
Index analysis
The study found that commercial voice cloning tools are accessible to the general public and capable of producing highly realistic voices that deceive most people, with 58 percent of participants failing to distinguish cloned voices from authentic ones.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.