People can no longer reliably tell AI and human voices apart, except in examples of conversational-style speech or with familiar voices, such as those of friends or family.
Reproducing human voices using artificial intelligence has long been a research goal, for example to help people who lose the ability to speak, but until recently, people have mostly been able to recognise an AI-generated voice. In 2023, researchers found that English and Mandarin speakers could differentiate between real and deepfake voices about 70 per cent of the time.
But at the University of California, Berkeley, and his colleagues have shown that the best AI voice generators are now much harder to spot. The findings will heighten fears about the risk of scams involving fake voices.
Advertisement
鈥淵ou think you鈥檙e good at [telling the difference], but you鈥檙e not,鈥� says Farid. 鈥淚 can鈥檛 tell you how many people I talk to who say to me, 鈥業 can tell the difference between the AI or not, it鈥檚 really easy.鈥� You鈥檙e wrong.鈥�
Farid and his team used AI company ElevenLabs鈥� voice cloning software to replicate a dataset of real human voices of 220 English speakers from the US. They included people of different races and genders, each responding to 32 different prompts, ranging from single-sentence replies of around a second to longer and unscripted answers of almost a minute.
Then, more than 600 people listened to different pairs of similar-sounding real voices and AI voices, without being told that AI was involved, before being asked whether they could tell if two voices were from the same source. The participants thought a real voice and an AI-generated cloned voice were the same about 80 per cent of the time.
When the researchers asked people to guess whether a voice was real or AI-generated, the volunteers answered correctly around 60 per cent of the time, or just above chance. However, people were better at spotting an AI voice when the speech sample was longer or less scripted compared with shorter, read responses.
鈥淲e鈥檙e at a point now where there are state-of-the-art deepfake voices that are sufficiently human in their qualities to be [indistinguishable] from genuine human recordings,鈥� says at University College London.
In a separate study, McGettigan and her colleagues also found that people struggled to differentiate between AI and human voices, doing roughly as well as a random guess. However, when listening to a voice they knew well, such as that of a family member, they were much better at detecting deepfakes.
Fake voices are likely to be even harder to detect in real-world situations compared with the controlled setting of an experiment, says Farid. 鈥淭his is the best possible situation. They鈥檙e sitting at home, they know they鈥檙e running an experiment, they鈥檙e paying attention, they鈥檙e getting paid for their time. It鈥檚 not a frantic call at 2 in the morning.鈥�
While there are statistical tools and AI-powered detectors that can help identify an AI voice from a recording, this is much harder to do with a live voice, so the best protection against AI voice scams is to agree on a password with your friends and family, says Farid. 鈥淵our family should have a secret passcode, and when somebody calls, you ask them what the passcode is.鈥�
Scientific Reports
Computers in Human Behavior: Artificial Humans
