Deepfakes: Learning How to Distinguish AI-Generated Faces from Real Ones
Image source, ANU (left; generated using StyleGAN3) / Igor Alecsander via Getty Images (right)
In 2024, a person working at an engineering firm named ‘Arup’ in Hong Kong joined a confidential acquisition meeting with colleagues via video call. Those in the financial sector attending saw nothing unusual at the time.
However, only after making a payment of 25 million US dollars did he realize he had been deceived.
While everything else seemed normal, the individual appearing as the chief financial officer on the video call was actually a ‘deepfake’—an AI-generated face.
Detecting AI-generated faces is becoming increasingly difficult every day.
According to consulting firm Deloitte’s 2024 report, approximately 40 billion dollars’ worth of fraud in the US is predicted to involve generative AI technology in the coming year.
This issue has prompted Amy Dowel, a psychologist and associate professor at the Australian National University, to question whether we can all learn to better distinguish AI-generated faces.
“At the moment, we are quite weak at detecting AI faces,” she said.
Our ability appears to be weaker than simple chance would suggest.
How skilled are you at recognizing real human faces versus AI-generated ones?
The ‘Super-Average’ Face
If you found the above test difficult, you’re not alone.
Dowel explains that difficulty in recognizing AI faces reveals a need for improvement in real-life identification. “The issues of scams with fake identities, deceit on dating sites, or fraud targeting bank accounts affect us all,” she said.
Together with colleagues from Australia, Canada, and Scotland, she developed a new training method that was published last month in the journal PNAS.
AI technology is continually advancing, and previously obvious errors have now been corrected by current systems, reducing the chances for scammers to exploit images with outdated flaws.
For this reason, researchers teach people to detect overall facial impressions rather than searching for small distortions like earrings or other minor details.
“The notion that AI-generated faces look strange is a misconception based on our previous experience,” Dowel said. “Real human faces contain subtle natural asymmetries and unique features, which currently sets them apart in unexpected ways.”
Image source, JohnnyGreig via Getty Images (left) / seb_ra via Getty Images (right)
Dowel’s prior research found that AI-generated faces were more frequently mistaken by people for real ones, a phenomenon she calls “AI hyperrealism.”
Since AI algorithms are influenced by mathematical averages derived from thousands of faces, AI-generated faces appear more average or “hyper-average,” Dowel explains.
This makes them appear more trustworthy and attractive than typical real faces.
The study included only white faces generated by AI because algorithms have been more extensively trained on such images.
Training
Researchers initially tested 45 adults on their ability to detect AI-generated faces. In one test, participants were shown three faces and asked to identify which was AI-generated, with an average accuracy of 41%.
Subsequently, they received about an hour of training using roughly 100 faces—half real and half AI-generated. Participants assessed each face based on distinctiveness, memorability, expression, symmetry, proportion, and attractiveness.
Dowel explains that real faces tend to be more distinctive, memorable, and expressive, while AI faces appear more uniform and attractive.
“We’re not saying AI faces are easier to detect, but the training teaches people how to evaluate the different features. The training points out which faces are AI-generated and which are real,” she said.
After training, correct identification rates rose to an average of 81%.
Similar training in Canada showed comparable improvements.
Dowel comments, “We observed similar strong effects there as well.”
Natural Tendencies
Dr. Alejandro Estudio, a psychologist at Bournemouth University in the UK, said the method used in the research is remarkable because it leverages humans’ natural thinking and comprehension skills.
“We distinguish faces by analyzing them holistically,” he explained.
For example, we don’t look at just a face’s distinctiveness or symmetry alone; we evaluate it overall.
“In daily life, practicing this natural holistic facial recognition process within education is likely the best approach,” he concluded.
Playing a ‘Cat-and-Mouse’ Game
The study has certain limitations.
According to Dowel, while this training is highly effective for the StyleGAN3 system, its effectiveness with other AI models remains untested.
“Your participants were young adults, so questions remain whether this training works for children or elderly people, who also face risks of scams,” she added.
People are currently up against increasingly complex AI systems, which is the major challenge, researchers note.
Though automated tools for detecting AI-generated content exist, they have limitations.
These tools work only on certain AI systems, and their efficacy may decrease as new models evolve. Therefore, human judgment remains central to this process.
“It’s like a cat-and-mouse game,” Dowel said. “Just as with computer viruses, these systems are constantly evolving and changing, and we must progress alongside them.”
Still, she notes the study’s findings provide grounds for optimism.
In the future, such training could be expanded to help identify not only AI-generated static faces, but also voices and videos. This would assist us in easily determining whether the person speaking on the phone or video call is real or an AI-generated deepfake.