AI faces seem realer than real

AI-generated faces are now so realistic that people often mistake them for being human. This particular problem is well-known — but new studies suggest that there's an even bigger one. 

Elizabeth J. Miller at the Australian National University and colleagues produce evidence in their new paper, published in Psychological Science, that White AI-generated faces are not only indistinguishable from human faces, but are perceived as being more human than real faces. This counterintuitive phenomenon is an example of 'hyperrealism'.

This new work follows three recent studies: one, from 2022, that found people were unable to distinguish AI from human faces, and two others that suggested people may over-identify AI faces as human. All three studies used the same StyleGAN2 algorithm for generating human faces, but varied in the races of the faces that they tested. There is a theory that especially average human faces may be perceived as being prototypical — and so be more likely to be judged as definitely human (or to be judged as being human, not AI, with a greater degree of confidence). Because StyleGAN2 was trained on mostly White faces, Miller and colleagues wondered if this might have biased the algorithm towards generating more 'especially average' White than 'especially average' non-White faces. 

To explore this, the team first re-analysed data from the first of those three studies. They found that White AI faces were judged as human significantly more often than White human faces, providing evidence of White AI face hyper-realism. In contrast, non-White AI faces were just as likely as non-White human faces to be judged as human.

Next, the team recruited 124 White adults, who were each shown 50 White AI faces and 50 White human faces of the same sex. (The team used the same race throughout to avoid any other-race effects.) The participants had to decide whether a face was AI or human, and indicate how confident they were in this judgement. They were also asked to give feedback on what information they used to make this decision. 

The researchers found that again, the White AI faces were judged as human more often than their human counterparts. In fact, the five faces judged as human most often were all AI-generated, while only one of the five faces most often judged as AI was actually AI. Unfortunately, results also suggested that participants who performed worst at detecting AI faces had the poorest insight into their abilities.

The team then looked at which attributes the participants reported as feeding into their judgements. There was a broad range of answers, but something about a specific facial feature (especially to do with the eyes), the presence or absence of wrinkles, judgements of facial symmetry, and the quality of the image (whether it appeared to be edited, for example, or its clarity) were among the more common replies. 

In a subsequent experiment, the team found that although AI faces were in fact more average (and therefore less distinctive) than real human faces, this didn't help the participants make accurate AI/human judgements. Instead, they relied on some other cues, such as memorability and familiarity, which in fact didn't help them — and also failed to use certain features, such as facial symmetry, that really could have informed accurate judgements. The participants did correctly use higher ratings of attractiveness, however, as being more suggestive of an AI than a human face.  

In a final element of the study, the team found that while humans are susceptible to AI hyper-realism, an AI detection algorithm could use data on the 14 attributes from their earlier study to accurately distinguish AI faces from human faces. 

This new work not only contributes to our understanding of what attributes we use to make judgements about faces, but also suggests that training AI algorithms on mainly White faces produces a disparity in how real they seem compared to non-White AI faces. This effect could easily impact any field in which AI faces are used as stimuli. This leads the authors to recommend that "studies using AI faces should verify that they are perceived as equally natural across races." They also recommend that everyone should be educated about the potential for AI hyper-realism.

Read the paper in full: https://doi.org/10.1177/09567976231207095