Dipfaces, synthetic media that mimic the voice or appearance of real people, pose a significant threat in the development of artificial intelligence. They can be used for manipulation and deception.
In order to assess how accurately people can recognize dipfaces, scientists from London conducted a study. They enlisted 529 participants to listen to audio recordings in English and Chinese, both real and fake, which were generated using a speech synthesis algorithm (TTS). The researchers created 50 samples of dipface speech in each language that did not overlap with the training data used by the algorithm. The participants were then tasked with identifying which samples were artificial.
The results, as published in a study recorded in the Journal PLOS ONE, indicate that the language of the dipface did not affect the accuracy of recognition. Even after providing the participants with examples of dipfaces to increase awareness, their ability to identify artificial content did not improve significantly.
The researchers concluded that individuals cannot reliably differentiate dipface speech, even with training that was meant to enhance their ability to distinguish artificial content from authentic content. The study also highlighted the increasing realism and difficulty in recognizing dipfaces as speech synthesis algorithms advance. Previously, creating a dipface required thousands of voice samples from a person, while now a mere three-second segment of speech is sufficient.
The challenge in detecting dipface speech underscores the potential for abuse, calling for the development of protective measures. A senior researcher at the University of Toronto’s Senior Citizen Lab, wrote that these current trends are crucial in combating fraud and distinguishing deceptive practices.
He further explained that skilled dipface creators and creative scammers may collaborate, quickly adapting to new situations. Telephone fraud and phishing schemes, in particular, rely on short feedback cycles where operators swiftly learn from their mistakes and successes.