People Can't Distinguish AI Voice Clones from Actual Humans Anymore (singularityhub.com)

A new PLOS One study led by Nadine Lavan (Queen Mary University of London) shows consumer-grade voice synthesis has become effectively indistinguishable from real humans. Researchers used ElevenLabs’ voice generator to produce 40 fully synthetic voices and 40 clones derived from a public dataset; each clone required roughly four minutes of recordings and minimal expertise or expense. Thirty participants rated voice realness and made binary human/AI judgments; while fully synthetic voices were judged somewhat less real, the clones matched real recordings closely and listeners could not reliably tell them apart. Participants also rated AI voices as more dominant and trustworthy in some cases. For the AI/ML community this is a milestone and a dilemma: state-of-the-art models can produce ultra-realistic speech with very little data, enabling scalable use in accessibility, education, and customer-facing agents but also lowering the barrier for impersonation, fraud, copyright infringement, and misinformation. The findings underscore the urgent need for technical defenses (watermarking, provenance, detection detectors), policy frameworks, and user-facing disclosure standards—because model-level guardrails alone may not keep pace with rapidly improving generative audio.