AI-generated audio that mimics humans can be so convincing that people can't tell the difference a quarter of the time – even when they're trained to identify faked voices, a new study claims.