15- to 20-Minute Training Helps People Spot AI-Generated Faces

Look at a polished face on a screen and the instinct can misfire: AI-generated faces—at least those of white people, the study says—can seem more human than real ones. At the University of Southampton, researchers tested DISCERN-AI, a 15- to 20-minute training session that moved participants from below-chance performance to above-chance performance.

That matters because synthetic faces are freely available online and already appear in romance scams, election interference, cyberbullying and espionage. A recent study found more than 7,000 accounts on X, formerly Twitter, using fake AI profile pictures to post spam. Mansi Pattni, co-lead author and a postgraduate researcher at Southampton, says people were more likely to select a hyperrealistic fake than a real face.

DISCERN-AI does not ask trainees to memorize a single tell. Its first section reverses instincts: proportionate, familiar faces can be clues that an image was generated by AI, while a memorable face can signal authenticity. The second highlights polished, high-quality images as more likely to be AI and distinctive, quirky images as more likely to be real. The final section tells trainees to set aside smooth skin and smiling, which the researchers say are not useful signals.

More than 600 participants were tested before and after training, against an untrained group, and in a surprise follow-up 20 days later. Tina Seabrooke, co-lead author of the paper, says the gains were not just greater skepticism: hit rates rose and false alarms fell, with the training also performing well compared with other misinformation treatments, such as spotting fake news. Computer models trained on the same cues achieved 94% accuracy, higher than the trainees.

So, concretely, a short lesson could give people a better first filter for profile pictures used in scams, election interference or cyberbullying, without relying on a vague feeling that a face looks strange. The immediate beneficiaries would be people making quick judgments online, but the ceiling is still visible: machines using the same cues did better, and the Southampton team only says it plans to make the training accessible online in the near future. The study points to a practical human skill, not human-level detection.