As Deepfake Scams Surge, Researchers Test Whether People Can be ‘Trained’ to Spot Fake Faces

The findings come as deepfake technology becomes more realistic and more widely used in scams, identity fraud, and misinformation campaigns.

Some of the faces generated using StyleGAN3 and used to train human detectors. Courtesy of Australian National University.

The rapid rise of deepfakes is making it increasingly difficult for people to tell what is real online, but new research from the Australian National University suggests humans can be trained to better detect AI-generated faces at a time when digital impersonation and fraud are escalating.

Researchers from the ANU Emotions and Faces Lab found that targeted training significantly improved people’s ability to distinguish between real and AI-generated faces, with some participants achieving near-perfect accuracy after instruction.

The findings come as deepfake technology becomes more realistic and more widely used in scams, identity fraud, and misinformation campaigns, with experts warning that visual “proof” of identity is becoming increasingly unreliable.

Deepfakes—synthetic media generated by artificial intelligence—are increasingly being used in financial scams, identity theft, and online impersonation. Recent research highlights that these technologies are now routinely deployed in fraudulent investment schemes, voice-clone scams, and fake celebrity endorsements designed to trick users into transferring money or sharing personal data.

Security analysts warn that the problem is no longer hypothetical.

Deepfake-enabled scams have been linked to billions of dollars in losses globally, with criminals using AI-generated video and audio to impersonate trusted individuals, including executives, public figures, and even family members.

According to the AI Incident Database, a crowd-sourced repository of media reports on AI misuse, reports of AI-related incidents rose 50 percent year-on-year from 2022 to 2024, and in the 10 months to October 2025, incidents had already surpassed the 2024 total.

At the same time, researchers note that automated detection tools are struggling to keep up. Many systems perform well in controlled environments but fail to reliably identify manipulated content in real-world conditions, particularly on social media platforms where most deepfakes circulate.

The result is what experts describe as a growing “trust gap”—where people can no longer assume that seeing or hearing someone online is proof of authenticity.

Training People to Spot What Machines Generate

Against this backdrop, the ANU study tested whether humans could be trained to recognise AI-generated faces more reliably.

Participants were taught to focus not on obvious visual flaws—such as distorted hands or irregular earrings—but on more subtle patterns in facial structure.

Lead researcher Associate Professor Amy Dawel said traditional “spot the glitch” approaches were becoming less effective as generative AI improved.

“AI faces tend to be more symmetrical, more proportional and more attractive, but without training we often interpret these features as signs of being human,” she said.

The training focused on six perceptual cues: distinctiveness, memorability, proportionality, symmetry, attractiveness and expressiveness.

After training, participants significantly improved their detection accuracy, with high performers reaching near-perfect results when identifying AI-generated faces produced by advanced systems such as StyleGAN.

“It was amazing to see the dramatic improvement in people’s ability to detect AI faces,” Dawel said.