New training method doubles accuracy in spotting AI-generated faces, study shows

Training People to Recognize AI Faces Shows Dramatic Results

Researchers at Australian National University have developed a training method that dramatically improves people's ability to spot AI-generated faces, with accuracy rates nearly doubling from 40% to 80% after just an hour of practice 1

1

. In a study published in PNAS, the Emotions and Faces Lab demonstrated that training people to recognize AI faces through broader perceptual patterns rather than fleeting visual glitches offers a human-centric solution to the deepfake problem 3

3

. Some participants achieved near-perfect accuracy, marking a significant advance in deepfake detection capabilities at a time when AI image generators produce portraits so convincing that even careful observers struggle to distinguish fact from fiction.

The Six Key Markers for Distinguishing AI Faces from Real Ones

The training method focuses on six perceptual qualities that differentiate synthetic media from authentic human faces: symmetry, proportionality, attractiveness, expressiveness, distinctiveness, and memorability 4

4

. "Our training directs people's attention to global qualities that differ between AI and human faces," said Amy Dawel, associate professor at Australian National University and lead author of the study 1

1

. Rather than instructing participants directly on what to look for, researchers showed them approximately 100 faces across six training blocks and asked them to rate each face on these qualities, allowing participants to discover the patterns through experience rather than direct instruction 3

3

.

Why AI Faces Drift Toward Statistical Averages

Current generative AI image generators train on datasets composed of millions of images, and when prompted to create a face, they compose new faces based on mathematical patterns shared across those datasets 1

1

. This reliance on "the mathematical average of the tens of thousands of faces on which they are trained" causes AI-generated faces to appear more typical than real human faces 3

3

. The result is faces that are more symmetrical faces, more proportional, and more attractive—while simultaneously being less expressive, less distinctive, and significantly less memorable 2

2

. This hyperaverageness creates a subtle banality that humans can implicitly sense, even if they don't consciously recognize it.

The Limitations of Previous Detection Methods

Past attempts to teach people to spot AI-generated faces focused on training viewers to look for visual glitches or statistical fingerprints left behind by particular image generators, such as a wonky ear or an eye with two pupils 1

1

. However, these clues can disappear with a software update or by simply using a different prompt. "The AI is getting too good," Dawel noted in a press release, adding that fraudsters may avoid using pictures with obvious flaws anyway 1

1

. Software-based deepfake detection tools also suffer from serious weaknesses—some can be fooled simply by converting the image type from png to jpg 2

2

.

The Growing Threat of Deepfake-Related Fraud

Deepfake faces generated via artificial intelligence have become so realistic that they routinely fool people, with research suggesting there may be $40 billion worth of deepfake-related fraud annually by 2027 2

2

. From 2023 to 2025, the volume of deepfakes online exploded with roughly 900% annual growth as AI-driven image generators improved 3

3

. Research from as early as 2023 discovered that some AI faces are "hyperreal"—they look more real than actual human faces—and people are overconfident they can detect AI slop imagery, with the most confident people making the most errors 2

2

. Apps like Zoom and Tinder now allow users to submit biometric identification, such as retinal scans, to help prove that a real person exists behind a profile picture 1

1

.

Training Effectiveness and Future Applications

The study involved 45 participants at Australian National University, with impressive results showing that all participants improved in their AI detection abilities 2

2

. "Even relatively short training sessions helped participants improve their accuracy," says Tanya George, a student researcher at Australian National University who trained the study's participants 1

1

. The Different Minds Lab at the University of Victoria in Canada conducted a replication that obtained results as strong as the original Australian study, demonstrating the training is reliable and can work for different groups of people 2

2

. The training was equally effective when administered online rather than in person, suggesting it could be a cost-effective remote intervention in deepfake detection 2

2

.

What Remains Unknown About This Approach

The training used faces produced with StyleGAN3, one of the most realistic face generators available, but the technology advances rapidly and many other models exist 2

2

. Researchers acknowledge that StyleGAN-generated faces may have different characteristics than those produced by other generative AI systems, and important questions remain: do the training benefits hold up over time, and is the training effective for people of all ages, including older adults or children 2

2

? "We've shown our training is effective for some of the most convincing fakes available, StyleGAN faces. Now we need to find out whether that training generalises to other AI-generated faces," Dawel stated 4

4

. The method has potential to adapt to new models by updating the training images and using multimedia, but evidence for this approach remains forthcoming.