Humans have a unique ability to recognise an individual human face in a fraction of a second. They automatically assess a stranger’s face in just 100 milliseconds. But even this evolutionarily refined ability is no match for the latest AI models, which can easily deceive it, according to the authors of a study published in the academic journal Journal of Vision.
“Our research shows just how much people are at risk of being deceived by AI-generated images,” said psychologist Alexis McGuire, who led the study. According to her, the problem is exacerbated by the fact that advanced artificial intelligence models are accessible to anyone without technical skills. Anyone with internet access can create fake faces that can easily be misused for various types of fraud. “It is important to inform the public about how easily such images can be created and how easily they can be misused — as well as the ways in which people can fall victim to them,” she stressed.
Can you feel jealous of artificial intelligence? According to Canadian research, the answer is yes (in Czech)
According to her, faces created by artificial intelligence are highly realistic and are becoming increasingly trustworthy. Newer and more sophisticated technologies generate fake portraits that can deceive people into thinking they are real in roughly one-third of cases.
In this experiment, the scientists presented 169 volunteers with a set of 96 highly diverse faces. They were asked to determine which were real and which had been created by AI. On average, the participants were successful in 58.4 percent of cases — which, as the authors note, is only slightly better than if they had been guessing at random. What surprised the scientists was the fact that more modern AI models were less successful at deceiving people than older ones.
The experts followed up this part with another experiment. They asked a different group of volunteers not to identify the source of the photograph this time, but to rate the trustworthiness of 96 randomly presented faces on a scale from 1, meaning very untrustworthy, to 7, meaning very trustworthy. Here, the results were clear: participants rated real faces as the least trustworthy, while faces produced by AI models appeared more trustworthy. The most advanced models, meanwhile, generated the most trustworthy-looking faces.
UN chief warns against unregulated AI (in Czech)
Scientists do not yet have an answer as to why faces created by artificial intelligence using a newer model were rated as less realistic than faces created by an older model, even though evaluators rated them as the most trustworthy. “It is a paradox that points to the possibility that judgments of realism and trustworthiness are driven by two distinct psychological mechanisms,” the authors suggest.
The authors warn that while this progress offers great opportunities, for example in film, these trustworthy but non-existent people may at the same time contribute to an overall erosion of trust.
“As AI-generated images become increasingly sophisticated and accessible, we as a society are increasingly exposed to artificially generated faces — often in malicious and exploitative scenarios such as political disinformation, financial scams, identity fraud or so-called ‘catfishing’. It is essential to understand the threat posed by this democratisation of generative artificial intelligence, while also developing strategies to mitigate potential harms to individuals, organisations and democracies,” McGuire concludes.
Link to the study (in English)
An article written by Tomáš Karlík (CT), initially published on 9 July 2026, 13:30 (CEST)