AI or Human? Majority of People Cannot Tell the Difference Between AI Generated and Human Faces

If you were shown photos of five real people and five AI generated faces would you be able to correctly sort them? New research says you are probably not.

collage of people
Would you have guessed this was generated by AI?

The usage and accuracy of artificial intelligence (AI) has grown exponentially in recent years. So much so that it is getting harder and harder to discern how much of the content we consume was generated by AI. The fast-paced rise of AI has come with limited research into the effects of the intelligence.

This study set forth to see if AI generated faces are detectable by people and even suggested that these faces look more human-like than real faces. Their findings are a frightening reminder that AI is getting extremely powerful.

Hyperrealism in fake faces

The Australian National University researchers suspected that AI has gotten so good at mimicking nature that people would favor the generated faces over natural ones. Their research used this idea as their inspiration.

Hyperrealism means that the generated picture looks incredibly lifelike, like a high-resolution photograph. However, hyperrealism is a depiction and is not actually a photograph of the real thing.

To see if their suspicions were true, researchers asked 124 adults to pick out AI faces from a mixed set of AI and real people. They then were asked their confidence in their choice. To build on previous research stating that AI is best at mimicking white faces, they used only white participants and faces.

The participants were fooled

Their suspicions ended up being true. The top three faces that were labeled as human by participants were actually AI generated. Inversely, the participants labeled three actual people as AI generated.

Not only were people bad at picking the fake faces, but they also did not know they were bad at it. 51% of participants were both poor at distinguishing faces and were confident that their answers were correct.

This poses great concern for the future of misinformation by AI. People potentially are being fooled, and they have no suspicion of it.

What threw them off?

The researchers did a second study of 610 participants to judge specific attributes of faces to determine if they were AI or real. They studied 14 attributes ranging from eye placement to image quality, to proportions.

This study found that the AI faces looked more average therefore were detected as real people. Faces with proportional and memorable features were more incorrectly identified as AI by the participants.

This is counterintuitive to what we might have thought before this study and shows the danger of AI. Peoples’ natural sorting practices are not accurate for AI faces now. People with malintent using AI generated faces could manipulate and trick people using this power.