Most of us have probably heard AI voices through personal assistants like Siri or Alexa, with their flat tone and mechanical delivery giving us the impression that we can easily distinguish between an AI-generated voice and a real person. But scientists now say that the average listener can no longer tell the difference between real people and “deepfake” voices.
See also: Meta: Conversations with AI chatbot will be used for personalized ads
In a new study published Sept. 24 in the journal PLoS One, researchers showed that when people hear human voices — alongside AI-generated versions of the same voices — they can't accurately identify which ones are real and which are fake.
The study showed that, while generic voices created from scratch were not considered realistic, AI voices trained on the voices of real people – deepfake audio – were found to be just as believable as their real-life counterparts.
The scientists gave study participants samples of 80 different voices (40 AI-generated voices and 40 real human voices) and asked them to mark which ones they thought were real and which were AI-generated. On average, only 41% of the AI-generated voices were initially misclassified as human, suggesting that it is still possible, in most cases, to distinguish them from real people.
See also: Major record labels close to major AI licensing deals
However, for AI voices cloned from humans, the majority (58%) were incorrectly classified as human. Only slightly more (62%) human voices were correctly classified as human, leading the researchers to conclude that there was no statistical difference in our ability to distinguish the voices of real people from their deepfake clones.
The results have potentially profound implications for ethics, copyright, and security. If criminals use artificial intelligence to clone your voice, it becomes much easier to bypass voice authentication protocols at your bank or trick your loved ones into transferring money.
See also: California: Governor Gavin Newsom signs AI law
Live AI voices can also be used to fabricate statements and interviews from politicians or celebrities. Fake audio can be used to discredit individuals or incite unrest, sowing social division and conflict. Scammers recently created an AI clone of the voice of Queensland Premier Steven Miles, using his profile to try to convince people to invest in a Bitcoin scam, for example.
