A new study revealed just how easily people can be tricked by AI-generated voices, with only one in 50 able to tell the difference.

A survey of over 1,000 residents in the US by cloud communications company, Ringover, has found that, initially, over three quarters (78.3%) of people were confident they could tell the difference between AI and human voices.

In reality, only 2% of those surveyed could correctly guess every AI voice tests given in the study, a dismal drop from their self-assured attitude.

Ringover conducted the survey in the US in response to Tom Hanks’ recent condemnation of a dental company using his voice and image through AI-generation. Deepfakes – the use of people’s images and voice to create pictures, video and audio with their likeness – has become a burgeoning issue thanks to recent AI advancements and has hit well-known celebrities in particular.

But even those out of the spotlight should be concerned by AI’s continued development in biometrics – AI-generated images and voices can be used to trump biometric security measures, putting people’s data, assets, and other personal information at risk.

The findings from Ringover’s survey are therefore concerning, showing just how accurate these AI-generated voices are becoming.

With around two-thirds (68.3%) of Americans beleiving their voice to be secure, it is important to check how easy it is to fake some of the most well-known voices in the country.

Guess Who -Celebrity Voice Edition 

The study tested Americans on five well-known celebrity voices in the country, including media mogul Oprah Winfrey, former President Barak Obama, pop start Miley Cyrus, comedian Kevin Hart, and the UK’s Prince Harry.

While only 2% identified the fake AI voice from the real celebrity voice in all five instances, over a third (35.5%) were able to tell the difference between at least one celebrity voice and their AI clone.

Oprah Winfrey was the most easily spotted, with 42.7% guessing her real voice correctly, while Prince Harry – the only one with a non-American accent – proved to be the trickiest with just 28.5% correctly identifying his voice.

The age group which identified the voices most accurately were Gen X, aged 45-54, which identified the AI clones 22.9% of the time, with the lowest age group being millenials (25-34) at 12.8%. While those over 65 struggled, their sample size was too low to have statistical significance in the study.

The accuracy across the voices may also have to do with popularity and familiarity about the different celebrities, which can range with age as well.

Recommended reading

• AI: Cyber-friend or Cyber-foe?
• Deepfake Tech Creates AI Hologram of Binance Exec
• Scot-Secure West | The AI Genie is Out of the Bottle

“This study highlights the importance of how easy it can be for us to be tricked by AI,” Renaud Charvet, co-founder and CEO at Ringover, commented.

“Voice recognition software could pose a huge potential risk to businesses, and customers, being able to bypass security on phone banking by mimicking other people’s voices in order to obtain personal details and access accounts is really concerning.

“We would advise that you check with your friends and family if you receive out of the ordinary phone calls from them. If you think you’re being scammed, reach out to the person to check if it’s really them, and then block the number if you suspect any suspicious activity.

“Whilst this study was conducted listening to celebrity voices, it shows how accurate these pieces of software are at mimicking the voices of other, and how misleading they can be.”