University of Southampton Researchers Develop 15-Minute Training to Help Spot AI Deepfakes

Researchers at the University of Southampton have created a targeted training module designed to significantly improve the human ability to detect synthetic, artificial intelligence-generated human faces. The findings, published in the academic journal Computers in Human Behavior, show that a brief 15- to 20-minute instructional session can measurably elevate a person's accuracy in distinguishing genuine human photographs from computer-generated deepfakes. Furthermore, empirical testing confirmed that the performance gains achieved through the program remained demonstrably effective 20 days after subjects underwent the initial instruction.

Rapid advances in artificial intelligence have allowed image generators to produce hyperrealistic depictions of human faces, particularly white facial portraits, that viewers routinely judge as more authentic than real human subjects. These freely available synthetic images pose growing security and societal challenges, finding widespread misuse in financial romance scams, covert election manipulation, digital cyberbullying campaigns, and state-sponsored espionage operations. A recent study referenced by the researchers highlighted that over 7,000 active accounts on the platform X, formerly known as Twitter, relied on AI-generated profile photos to publish automated spam content across the network.

Without specialized instruction, untrained individuals frequently struggle to identify synthetic portraiture, performing below baseline statistical expectations. "People are surprisingly bad at spotting these hyperrealistic images," said Mansi Pattni, co-lead author of the study and a postgraduate researcher at the University of Southampton. "We perform worse than if we were to flip a coin and are more likely to choose fake hyperrealistic faces over real people."

To counteract these perceptive biases, the academic team engineered a structured, three-part training methodology named DISCERN-AI. The first segment of the program targets deceptive visual assumptions that observers naturally rely upon when examining imagery. Viewers typically perceive faces with highly proportional and familiar structures as authentic human beings, even though these characteristics are common indicators of algorithmic synthesis. Conversely, individuals often assume that a highly memorable or striking face was created by artificial intelligence, despite high memorability actually serving as a hallmark of genuine human photography.

The second module of DISCERN-AI conditions trainees to focus on useful visual signals that are commonly overlooked during routine inspection. Highly polished, flawless, and ultra-high-quality portraits carry a higher statistical probability of being AI-generated, while photographs containing distinct, quirky, or non-standard visual features are more likely to depict real people. The concluding phase of the program teaches users to disregard irrelevant features—such as smooth skin textures or facial expressions like smiling—that individuals often mistake for red flags, despite their lack of predictive value in identifying deepfakes.

The researchers verified the tool's real-world efficacy through extensive testing involving more than 600 study participants evaluated under diverse conditions. The experimental protocols evaluated cohorts before and after completing the coursework, compared trained individuals against uninstructed control groups, and incorporated a surprise re-test administered 20 days post-training to measure long-term cognitive retention. Across all scenarios, the data showed consistent improvements in visual discrimination skills.

"The results consistently showed that DISCERN-AI was effective in moving people from a below-chance performance to an above-chance performance," noted Dr. Tina Seabrooke, co-lead author on the publication and researcher at the University of Southampton, as first reported by TechXplore. "The training didn't just make people more skeptical, it improved accuracy, increasing hit rates and reducing false alarms. It also performed well compared to other misinformation treatments, such as spotting fake news."

While human trainees demonstrated substantial performance gains following the short educational intervention, automated computer systems trained on the exact same diagnostic cues achieved even higher levels of precision. Automated machine learning models incorporating the DISCERN-AI criteria successfully identified synthetic portraits with a 94% accuracy rate. The University of Southampton research team plans to make the DISCERN-AI training system freely accessible online to the general public in the near future.