AI-Generated Faces Fool Even Super Recognizers, But Five-Minute Training Boosts Detection

AI-Generated Faces Deceive Even Elite Facial Recognition Experts

Artificial intelligence has reached a troubling milestone in visual realism. AI-generated faces now fool even super recognizers—an elite group with exceptionally strong human facial recognition abilities—who perform no better than random chance when attempting to identify AI-generated faces 1

1

. According to research published in November in the journal Royal Society Open Science, super recognizers correctly identified fake faces only 41% of the time, while people with typical recognition capabilities fared even worse at just 31% accuracy 2

2

. This means most people are actually worse than chance, mistaking synthetic faces for real ones more often than not.

The study, led by Katie Gray, an associate professor in psychology at the University of Reading in the U.K., tested 664 participants on their ability to distinguish real human faces from images generated by StyleGAN3, an advanced AI software 2

2

. The increasing realism of AI-generated images has created what researchers call "hyperrealism"—a phenomenon where individuals are duped into thinking fake faces are more real than actual human faces 1

1

.

Five-Minute Training Session Dramatically Improves Detection Accuracy

Despite the alarming baseline results, the research offers encouraging news. A brief five-minute training session on common AI rendering errors significantly improved participants' ability to identify AI-generated faces 1

1

. After training, super recognizers achieved 64% accuracy in fake face detection, while typical recognizers reached 51%—representing improved detection accuracy of 23 and 20 percentage points respectively 2

2

.

The training highlighted specific telltale signs in AI-generated faces, including unnatural hair patterns, odd-looking hairline errors, unnatural skin texture, having a middle tooth, and faces that appear more proportional than real ones 3

3

. Participants were shown 10 faces during training and received real-time feedback on their accuracy, followed by a recap of rendering errors to watch for 3

3

.

"I think it was encouraging that our kind of quite short training procedure increased performance in both groups quite a lot," Gray told Live Science 1

1

. Trained participants also took longer to scrutinize images—typical recognizers slowed by about 1.9 seconds and super recognizers by 1.2 seconds—suggesting that careful examination is key to distinguishing AI from reality 1

1

.

Real-World Security Risks Demand Better Detection Methods

The implications extend far beyond academic curiosity. Computer-generated faces pose genuine real-world security risks, having been used to create fake social media profiles, bypass identity verification systems, and produce false documents 2

2

. Recently, TikTok users exposed deepfake doctors scamming social media users with unfounded medical advice 3

3

.

These deepfake faces are created using Generative Adversarial Networks (GANs), a two-stage algorithm where a fake image is generated based on real-world images, then scrutinized by a discriminator that determines whether it's real or fake 1

1

. Through iteration, these algorithms have improved to the point where synthetic faces routinely pass as authentic.

Gray suggests combining training with super recognizers' natural abilities could help tackle these challenges. "To best detect synthetic faces, it may be possible to use AI detection algorithms with a human-in-the-loop approach—where that human is a trained SR [super recognizer]," the authors wrote 1

1

. The super recognizers in the study were recruited from the Greenwich Face and Voice Recognition Laboratory volunteer database, having performed in the top 2% of individuals in facial recognition tasks 1

1

.

One caveat remains: participants were tested immediately after training, leaving questions about long-term retention 3

3

. Still, equipping people with tools to spot AI-generated content is essential as synthetic media floods social platforms. The challenge isn't limited to visual content either—researchers recently claimed ChatGPT passed the Turing Test, meaning AI language models are also becoming indistinguishable from humans 3

3

.