Southampton training programme helps people spot AI-generated faces

University of Southampton researchers have built DISCERN-AI, a 15–20 minute training programme that measurably improves people’s ability to tell AI-generated “hyper-realistic” faces from real human ones — with effects lasting at least 20 days.

The study, published in the journal Computers in Human Behaviour, addresses a problem the researchers describe as urgent: AI-generated faces — of white people at least — have become hyper-realistic, meaning people rate them as more real than actual photographs. Such synthetic faces are freely available online and have already been used in romance scams, election interference, cyberbullying and espionage; one recent study found more than seven thousand X accounts using fake AI profile pictures to post spam.

How the training works

Co-lead author Mansi Pattni says people perform worse than a coin flip at spotting these images and tend to choose fake hyper-realistic faces over real ones. DISCERN-AI attacks that in three parts. The first re-trains instinctive but misleading cues: we judge proportionate, familiar-looking faces as human when those traits can signal AI generation, while treating memorable faces as fake when memorability actually suggests authenticity. The second part teaches useful but overlooked signals, such as perfectly polished, high-quality images being more likely AI while distinctive, quirky ones tend to be real. The third tells trainees to ignore supposed red flags — smooth skin or a smile — that do not actually help.

The results

More than 600 participants were tested across controlled scenarios: before-and-after training, trained versus untrained groups, and a surprise follow-up 20 days later. Co-lead author Dr Tina Seabrooke says the training consistently moved people from below-chance to above-chance performance, increasing hit rates and reducing false alarms rather than merely making people more sceptical. Computer models trained on the same cues reached 94% accuracy — better than the human trainees. The team plans to make the training freely accessible online, and the research was funded by Southampton’s Web Science Institute.

Our opinion

The quiet headline here is that the cues exist at all. Humans are not doomed to lose this game; we are simply tuned to the wrong signals, and a short course can retune us. The 20-day persistence matters more than the immediate gain, because deepfake exposure is a drip, not a single event. The 94% machine result is a useful rebuke to pure fatalism, but also a trap: automated detectors become a new arms race while trained scepticism generalises. The honest framing is that DISCERN-AI buys back some ground for humans — it does not win the war.