Most people are overconfident in their ability to identify faces generated by AI, according to a new study by researchers at UNSW Sydney and the Australian National University (ANU).
The research, published in the British Journal of Psychology, involved 125 participants, including 36 people deemed to be ‘super recognisers’ and 89 control participants.
Super recognisers, comprising 1–2 per cent of the population, have exceptional memory for faces – meaning they can recognise people they met only briefly years earlier, pick out familiar faces even after drastic changes to their appearance, and spot background actors in television shows and films that most people would miss.
In an online test, the super recognisers and control participants had to judge whether a series of faces were real or AI-generated.
“We wanted to know whether super-recognisers are also better at spotting fake, AI-generated faces,” study author and UNSW School of Psychology researcher Dr James Dunn told BBC Science Focus.
The answer? Yes, they were – but only marginally better than the controls, who themselves performed only slightly better than chance. Specifically, control participants scored an average of 50.7 per cent correctly, while super recognisers had an average score of 57.3 per cent.
The researchers were surprised by the small impact of being a super recogniser on participants’ abilities to spot an AI fake.
In fact, they found that some control participants outperformed the super recognisers, suggesting there might be ‘super-AI-face-detectors’ out there with specific skills for detecting artificial faces.
What was consistent across all participants, however, was their confidence in their abilities – even when it wasn’t matched by their performance.
The researchers warn that this overconfidence could make people more vulnerable to scams as well as fabricated identities on social media, dating sites and professional networking.
While AI-generated images of people used to feature bizarre (and hilarious) distortions – extra hands, glasses fused onto faces, mismatched backgrounds – the latest technology has made them near-indistinguishable from real images.
So how can you improve your skills at recognising AI?
“Ironically, the most advanced AI faces aren’t given away by what’s wrong with them, but by what’s too right,” said study author and ANU psychologist Dr Amy Dawel. “Rather than obvious glitches, they tend to be unusually average – highly symmetrical, well-proportioned and statistically typical.
“It’s almost as if they’re too good to be true.”
Think you can do better? You can take a demo of the study test here.
Answers to above image: faces 2, 3, 5, 8, 9, 11 are AI-generated.
Read more:
