Researchers discovered that images of faces created by artificial intelligence (AI) are perceived as more trustworthy than photos of real faces, raising concerns about the risks of digital fraud and other damages.
First study using diffusion technology
This investigation is pioneering in analyzing the credibility of AI faces produced using the latest diffusion technology. The research was led by Alexis McGuire, along with Paul Taylor and Sophie Nightingale from Lancaster University; Maty Bohacek from Stanford University; and Hany Farid from the University of California, Berkeley.
Risks of deception by AI images
Alexis McGuire, a PhD student in Psychology, warned Phys.org that the research demonstrates that individuals are susceptible to being misled by AI-generated images. She explained that these AI models have made the online environment more accessible, allowing anyone to create fake faces without technical knowledge, which can be used for various types of harm.
McGuire emphasized the importance of educating the public about the ease of creating these images and their inappropriate uses, citing examples such as the spread of fake news, identity fraud, and catfishing. While humans are skilled at processing authentic faces, evaluating them in just 100 milliseconds, synthetic AI faces are becoming extremely realistic and more trustworthy, with new technologies managing to deceive people approximately one-third of the time.
Details of the experiments conducted
In an initial test, 169 participants observed a collection of 96 faces (varying in ethnicity, sex, and age) presented randomly, having to determine whether each was real or synthesized by AI. The average accuracy achieved by the participants was 58.4%, a result only marginally above chance (which would be 50%).
Interestingly, faces generated by the most current AI diffusion model (DM) were classified as less realistic compared to those created by a previous AI model (GAN). In a subsequent experiment, a new group evaluated the credibility of 96 faces on a scale of one (very untrustworthy) to seven (very trustworthy).
Real faces received the lowest confidence score, with an average of 4.03. However, both types of synthetic AI faces were judged as more trustworthy than real faces. Specifically, faces produced by the diffusion model (DM) were considered so trustworthy that they surpassed both real faces and GAN faces. GAN faces achieved an average confidence of 4.36, while diffusion-synthesized faces (DM) were the most trustworthy, with an average of 4.70.
The psychological paradox discovered
The researchers find it intriguing that faces created by the latest DM model were perceived as less realistic than those from the previous GAN model, yet were still considered the most trustworthy. McGuire pointed out that this finding reveals a paradox and suggests that the criteria for realism and trustworthiness may be influenced by two distinct psychological processes.
She warned that AI faces generated by DM technology could lead to a general deterioration of social trust. According to her, as AI images become more advanced and accessible, society becomes increasingly exposed to artificial faces, often in harmful and exploitative contexts, such as political disinformation, financial and identity fraud, and catfishing. Therefore, it is crucial to understand the threat that the democratization of generative AI represents and to develop methods to mitigate potential harm to individuals, organizations, and democracies.
Dissemination and invitation to participate
The research was published in the Journal of Vision under the title “AI-Generated Faces are Becoming More Trustworthy”. Anyone interested can participate in an anonymous online study called “Examining Individual Differences in the Detection of Real and AI-generated Faces,” where they will be asked to evaluate faces as real or AI-generated, in addition to answering other questions about their level of trust, culminating in a final score.