A 15-to-20-minute training session built at the University of Southampton moved more than 600 participants from worse-than-chance to better-than-chance accuracy at telling real human faces from AI-generated ones, and a surprise retest 20 days later found the gain intact. The course, DISCERN-AI, comes in three parts.
Image generators have come far enough that AI faces of white people look hyper-realistic, and observers rate them as looking more genuine than the faces of real humans. Such faces cost nothing, and already turn up in romance fraud, election meddling, online bullying and spying. Separate research identified more than 7,000 accounts on X pushing out spam from behind bogus AI-made profile photos.
The three segments work by inverting the cues people trust. The first takes on instincts that mislead. A face read as well-proportioned and familiar strikes viewers as more human, when those very qualities can be signs it was generated; a face people find memorable, the kind they are inclined to suspect, is actually more likely to be real.
The second covers cues that get overlooked but work: a flawlessly polished, high-quality picture is the more likely AI product, an odd and distinctive one the more likely photograph. The third tells trainees to ignore the warning signs they commonly assume, such as smooth skin and a smile, because those features are of no real help in separating genuine faces from generated ones.
The work, published in Computers in Human Behavior, tested the tool across several designs: performance before and after training, a trained group set against an untrained one, and the 20-day follow-up.
It did more than make people sceptical, says Tina Seabrooke, a co-lead author. Trainees correctly identified images more often and wrongly flagged them less often, and the course performed well against other anti-misinformation training, such as courses on spotting fake news.
Even so, trained people remained the weaker detector: software given the same cues outscored them, at 94% accuracy.
The course is not available yet; the team plans to put it online for anyone to use.