Study Finds People Trust AI-Generated Faces More Than Real Ones

Can individuals reliably tell the difference between genuine human faces and those created by artificial intelligence? A recent study conducted by researchers at Lancaster University indicates that this task is far more challenging than expected. Their findings show that people’s ability to differentiate AI-generated faces from real ones does not exceed random chance. Moreover, the study reveals that people are more inclined to trust faces produced by AI than actual human faces.

The British newspaper Daily Mail reported that the researchers warned this situation poses a significant risk, as it could enable identity theft or online fraud using AI-generated images. Alexis McGuire, the lead researcher and a doctoral candidate at Lancaster University, stated, "The tendency of people to perceive AI-generated faces as trustworthy makes them highly effective tools for fraud and online deception. For instance, a text-based scam could become more convincing if paired with a face that people instinctively trust."

Previously, distinguishing AI-generated faces was relatively easier due to noticeable flaws such as extra fingers, irregular teeth, or asymmetrical ears. However, recent research suggests that these indicators no longer improve people’s ability to detect deepfakes. Fraudsters can now easily correct or avoid such errors, making detection more difficult.

Challenges in Detecting AI-Generated Faces

The study highlights that the latest generation of image-generating models has become nearly impossible for humans to identify. Published in the journal Vision, the research involved 169 participants who evaluated 96 faces—both real and AI-generated. Each participant was shown a randomly selected face and asked to determine whether it was AI-created or authentic.

On average, participants achieved only 58.4% accuracy, a rate comparable to guessing by chance. Accuracy varied depending on the ethnicity of the faces and the specific AI model used, but the overall trend remained consistent. Real human faces consistently received the lowest trust ratings, averaging 4.04 on a scale from 1 to 7, where 7 indicates very high trustworthiness.

In contrast, AI-generated faces, which were difficult to distinguish, received higher trust scores, averaging 4.36. Faces produced by diffusion models were rated as the most trustworthy, with an average score of 4.7. Interestingly, this suggests that people trusted diffusion model-generated faces more, even though they believed these faces appeared less realistic.