A growing body of research suggests that artificial intelligence is making it increasingly difficult for people to distinguish between real and synthetic speech, with most individuals unable to reliably tell the difference—even after receiving training.
A team of Chinese researchers from Tianjin University and Chinese University of Hong Kong conducted an experiment involving 30 participants, who were connected to brain-scanning devices while listening to a series of audio recordings. The participants were asked to determine whether the clips were generated by artificial intelligence or were genuine human voices.
In most cases, participants failed to make an accurate distinction, with the research team describing the group as “weak in discriminating between the two types.”
The researchers then attempted to train participants to improve their ability to detect synthetic voices. However, the results showed only marginal improvement, suggesting that even targeted training may have limited impact on enhancing human detection capabilities.
Despite this, the study revealed a potentially important development at the neurological level. According to the researchers, training appeared to make the brain’s responses more distinct when processing human speech compared to AI-generated audio.
Xiang Bin Ting, who led the research, explained that “the auditory system in the brain seems to begin detecting subtle acoustic differences, even if individuals are unable to translate these signals into reliable behavioural decisions.” He added that these faint recognition signals are nonetheless “encouraging” for future research.
The findings follow earlier research published in September 2025 by Queen Mary University of London, which warned that AI-generated “deepfake” voices created using widely available software have become virtually indistinguishable from real human speech.
Evidence from related studies suggests that people perform only slightly better when identifying AI-generated images. A study published in February 2026 by University of New South Wales and Australian National University found that most individuals tend to overestimate their ability to detect synthetic faces.
Concerns over the growing sophistication of such technologies have also been raised by the financial sector. In 2025, Citibank issued a warning about increasingly convincing AI-generated audio and visual deepfakes, highlighting their spread across areas such as recruitment processes, financial transactions, and executive impersonation.
Taken together, these findings underscore the mounting challenges posed by AI-generated content, particularly as advances in deepfake technology continue to blur the line between reality and fabrication.
