Researchers help people detect AI faces

Researchers at the University of Southampton have come up with a way to help people spot AI-generated faces.

They’ve developed a short training programme that improves people’s ability to distinguish between real, human faces and those generated by AI software.

A study, published in the journal Computers in Human Behaviour , shows that the training was still effective 20 days later.

AI has developed to the point where the ‘deepfake’ faces it generates (of white people at least) are ‘hyper realistic’, meaning people find them to be more real than actual human faces.

Freely available online, these synthetic faces are already being used in romance scams, election interference, cyberbullying, and even espionage. A recent study revealed more than seven thousand accounts on X (formerly Twitter) were using fake AI profile pictures to post spam content.

“People are surprisingly bad at spotting these hyper-realistic images,” says Mansi Pattni , co-lead author and 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 hyper-realistic faces over real people.”

To correct this, the team at the University of Southampton developed DISCERN-AI. It’s a fifteen to twenty minute training session consisting of three parts.

The first focuses on features we rely on but that deceive us. We instinctively judge faces that are proportionate and familiar as more human, but these are actually clues it could be AI generated. Conversely, we think a memorable face is more likely to be AI, when in fact it’s a sign of authenticity.

Part two trains people on things that are useful but often overlooked. Perfectly polished, high-quality images are more likely to be AI, while distinctive, quirky images are more likely to be real.

The last section asks trainees to ignore what we think are red flags - such as smooth skin or smiling. These aren’t actually helpful in discerning between real and AI-generated faces.

To test the effectiveness of the tool, researchers tested over 600 participants in different scenarios, such as before and after training, a group receiving the training versus one that didn’t, and a surprise follow-up test 20 days after the training.

“The results consistently showed that DISCERN-AI was effective in moving people from a below-chance performance to an above-chance performance,” says Dr Tina Seabrooke , co-lead author on the paper, also from the University of Southampton.

“The training didn’t just make people more sceptical, 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 the research shows humans can be trained to spot AI faces more effectively, computer models trained on the same cues performed even better than the trainees, achieving 94% accuracy.

The team has plans to make the training accessible to people online in the near future.

The research was funded by the University of Southampton’s Web Science Institute .