New Study Reveals Human Difficulty in Detecting Convincing Audio Deepfakes, Raising Concerns of Increased Fraud and Scams
A recent investigation conducted by University College London has uncovered that AI-generated audio deepfakes are incredibly convincing, making it challenging for humans to distinguish between genuine and manipulated voices. The prevalence of deepfakes, a form of AI-generated content, has resulted in growing worries regarding security threats and potential misuse.
In order to carry out the research, the team employed a text-to-speech algorithm to produce 50 deepfake speech samples in both English and Mandarin. Participants were then asked to determine whether the speech originated from a real human or a computer-generated source after listening to these samples. Unexpectedly, the accuracy rate for correctly identifying deepfake speech was a mere 73%, even among participants who had undergone training to recognize artificial content.
Lead author Kimberly Mai emphasizes the potential dangers associated with speech deepfakes, stating that humans are consistently incapable of effectively detecting them, even with training. Additionally, the study underscores the continuous evolution of deepfake technology, which could potentially make identification even more challenging in the future.
The study also found that both English and Mandarin speakers encountered similar difficulties in detecting deepfakes, emphasizing the universal threat potential across different languages.
Instances of audio deepfake misuse have already been observed in various criminal activities. In 2019, a UK-based energy firm fell victim to a scam in which an AI-generated voice impersonated the company’s CEO, resulting in significant financial loss.
As voice cloning scams continue to rise, the study cautions that identifying AI-generated speech will remain a formidable task. The Federal Trade Commission has issued a warning regarding scammers increasingly relying on voice cloning technology to defraud unsuspecting victims.
To combat this escalating problem, the study proposes exploring alternative approaches to mitigate the risks associated with audio deepfakes. Implementing regulations and policies, enhancing automated detectors, and utilizing crowd-sourced fact-checking measures are among the strategies that could be effective in addressing this issue. As technology continues to advance, finding dependable solutions becomes imperative in safeguarding against AI voice-cloning scams and protecting individuals from potential harm.
