Deepfakes: How to Learn to Identify AI-Generated Faces

In 2024, an individual working at an engineering firm called ‘Arp’ in Hong Kong participated in a “video call” with colleagues about confidential acquisition discussions. Being from the financial sector, he found nothing unusual at the time. However, after making a payment of 25 million US dollars, he realized he had been scammed. Although everything else appeared normal, the person seen in the video call, who resembled a chief financial officer, was actually an AI-generated ‘deepfake’ face.

Identifying faces created by AI is becoming increasingly challenging every day. According to a 2024 report by the consulting firm Deloitte, fraud using ‘generative AI’ technology could amount to nearly 40 billion dollars in the United States alone within the coming year. This has prompted Amy Dowel, a psychologist and associate professor at Australian National University, to ask—can we all learn to better recognize AI-generated faces?

“Currently, we are somewhat weak at detecting AI-generated faces,” she explained. According to Dowel, if AI-generated faces are hard to identify, real-world improvements are necessary. “Issues such as identity fraud, dating site deception, or banking scams affect us all,” she said. Collaborating with colleagues from Australia, Canada, and Scotland, she helped develop a new training method, which was published last month in the journal PNAS.

Dowel stated, “We are not claiming to be perfect at spotting AI faces, but our approach teaches people to evaluate based on various features.” After training to distinguish between AI-generated and real faces, accuracy rates averaged 81 percent.

Dr. Alejandro Estudio remarked that the method used in this research is remarkable because it leverages humans’ natural thinking and comprehension abilities.

According to Dowel, this training is highly effective for the StyleGAN3 system but still needs to be tested on other AI models. “People are currently competing against increasingly complex AI systems, and that is the main challenge,” she added.

Although automated tools to detect AI-generated content are available, they have limitations.

Dowel likened this to a game of cat and mouse, explaining, “Like computer viruses, these systems constantly evolve and change.”

Nevertheless, she said the study’s findings provide sufficient grounds for optimism.