A recent study from UCL reveals that humans correctly identify artificially generated speech only 73% of the time, regardless of whether the language is English or Mandarin.
The research raises concerns about the rapidly advancing capabilities of deepfake technologies and their potential misuse.
Deepfakes and Generative AI
Deepfakes are synthetic media crafted to mimic the appearance or voice of a real person.
This technology is a subset of generative AI, which utilizes machine learning to recreate sound or imagery based on large data sets.
Earlier algorithms needed thousands of voice samples, but current ones can recreate voices from as little as a three-second clip.
User-Friendly Tools Emerging
Apple recently introduced a tool that lets users reproduce their voice with just 15 minutes of recordings.
The UCL team used a text-to-speech algorithm trained on two separate data sets (in English and Mandarin).
Out of 529 participants, only 73% could correctly differentiate between real and deepfake speech, and their performance improved only slightly after receiving special training.
Implications and Concerns
Kimberly Mai, the lead author, highlighted that even with training, humans have difficulty reliably detecting deepfakes. Given that the research utilized older algorithms, more advanced deepfake tools might be even harder to discern.
Although generative AI offers benefits, such as aiding those with speech impairments, its misuse can be detrimental.
A notable instance from 2019 saw a CEO being deceived into transferring significant funds due to a deepfake of his superior’s voice.
Future Directions and Recommendations:
The researchers are aiming to develop improved automated detectors to identify deepfake audio and images.
Professor Lewis Griffin emphasized the need for proactive strategies by governments and institutions to manage potential misuse, while also acknowledging the potential benefits of the technology.
While deepfake technologies offer promise in areas like accessibility and communication, they also present significant challenges in security and misinformation.
As these tools become more accessible and sophisticated, it’s essential to strike a balance between harnessing their potential and mitigating their risks.
The study was published in PLOS ONE.
