People can only correctly identify deepfake speech 73% of the time, according to a new study. — Photo: Christin Klose/dpa
DUBLIN: More than a quarter of so-called "deepfake" voices fool even the sharpest listeners, going by a survey by University College London (UCL).
More than 500 people could only correctly identify deepfake speech 73% of the time, according to the UCL researchers, who trained "some" participants to prick up their ears for bot voices.
The survey was done in English and Mandarin Chinese, with much the same detection result across both languages, though the UCL team found English speakers "more often referenced breathing, while Mandarin speakers more often referenced cadence, pacing between words, and fluency" when identifying deepfake speech.
deepfake voices have already been used to con money from people who believed they were talking to a friend or business partner, with recent advances in artificial intelligence (AI) fuelling concerns that misuse of the technology could be accelerated as deepfake voices are made to sound more authentic.
"Due to technological advances, it is possible to produce a realistic-sounding clone using only a few audio samples," the UCL team warned.
The researchers, whose findings were published by the journal PLOS One, said the results likely did not reflect real-world scenarios, given even the untrained participants were aware they were taking part in a survey and therefore primed to listen up for deepfakes.
Most efforts at combating deepfake audios have so far used machine-learning detectors, which perform at about the same level as the participants in the UCL survey and better when it came to "unknown conditions for which automated detectors may not have been directly trained." But since deepfakes are likely to be improved, the best response would be to develop more sophisticated machine detectors, the researchers concluded.
Survey results from the US published in June found people to be "too confident" of their ability to identify deepfake videos. The same month, academics in Britain and Ireland warned that AI advances could result in deepfake videos and audio of dead people. – dpa
