Fake AI faces are more convincing than real ones

Face images created byartificial intelligence (AI) are now capable of inspiring more trust than images of real faces. This is what emerges from a new study led byLancaster University and published on Journal of Vision.

First author Alexis McGuire explained: “Our research shows that people are at risk of being deceived by AI-generated images. These AI models have democratized the online space and are accessible to anyone, even without technical skills, to create fake faces that can be used for a variety of malicious purposes. It is important to inform the public about the ease with which such images can be created, the potential for abuse and the ways in which they could become victims, for example through the spread of misinformation, identity theft and catfishing”.

How to recognize a fake face

Humans are experts at processing real faces, managing to automatically evaluate one in as little as 100 milliseconds. However, faces generated by artificial intelligence are extremely realistic and are becoming increasingly reliable thanks to newer and more sophisticated technologies, capable of creating fake images that deceive people, making them believe they are authentic in about a third of cases.

When 169 participants were asked to look at a collection of 96 faces (differing in ethnicity, gender and age) presented in random order and to indicate whether each face was real or synthesized by AI, their average accuracy was 58.4 percent, only slightly higher than that of a random choice (similar to flipping a coin, with a 50 percent chance).

The paradox of the latest models

Surprisingly, the faces generated by the latest AI diffusion model (DM) were judged to be less realistic than faces produced by a previous AI model (HOWEVER).

However, in a subsequent experiment, a new group of participants was asked to rate the trustworthiness of 96 randomly presented faces on a scale from one (not at all trustworthy) to seven (very trustworthy). Real faces were rated the least trustworthy, with an average trustworthiness score of 4.03. Both types of AI-synthesized faces were rated as more reliable than real faces, while faces produced by the diffusion model (DM) were found to be more reliable than both real faces and those generated by GANs. Faces generated by GANs received an average reliability score of 4.36, while faces synthesized via diffusion modeling (DM) were the most reliable, with an average score of 4.70.

According to the researchers, it is disconcerting that AI-synthesized faces generated by the latest AI diffusion model (DM) were rated as less realistic than faces produced by an older model (GAN), despite still being rated the most reliable.

McGuire said: “This result presents a paradox and therefore highlights the possibility that judgments of realism and reliability are driven by two different psychological mechanisms.”

The risks for society

He warned how faces generated by artificial intelligence using the latest digitization technologies could contribute to an overall erosion of trust in society. “As AI-generated images become more sophisticated and accessible, our society is increasingly exposed to artificially created faces, often in nefarious and exploitative contexts, such as political disinformation, financial and identity fraud, and catfishing. It is critical that we understand the threat this democratization of generative AI poses, as well as develop strategies to mitigate the potential harm to individuals, organizations, and democracies.”