Researchers found that people are generally bad at identifying which faces are real and which are generated by AI. Hyperrealistic AI faces are more often judged as real than actual human faces.
The study, led by the Australian National University, included experiments to see whether people could differentiate between human and AI-generated faces and which features led to the choices people made.
Previous studies showed that participants could spot a non-White AI face around 50% of the time. In this more recent study, a group of 124 adults were shown a mix of 100 AI and 100 real White faces and asked to identify if they were real or not.
The AI-generated faces were misidentified as human 65.9% of the time while human faces were judged as human only 51.1% of the time.
Co-author Dr. Eva Krumhuber said, “Artificial intelligence has reached an astonishing level of realism, and here we find that sometimes it can even seem more real than reality – hyperrealism – so that we can be very easily tricked into thinking an AI-generated face is real.”
The participants were also asked how certain they were of their evaluation. In a classic example of the Dunning-Kruger effect, the participants who made the most errors were the most confident that they were making correct judgments.
Elizabeth Miller, study co-author and PhD candidate at Australian National University said, “Concerningly, people who thought that the AI faces were real most often were paradoxically the most confident their judgments were correct.”
How would you have fared in the test? Which of these are real and which are AI-generated? We’ll let you know how you did lower down.
The researchers concluded that AI-generated faces aren’t just indistinguishable from human faces, but they have features that make them seem even more real to us. They termed this feature of AI faces hyperrealism.
The researchers used StyleGAN2, a generative adversarial network (GAN), to generate the AI faces. StyleGAN2 was trained on a large dataset of human faces with around 69% being White and 31% for all other races combined.
The researchers concluded that the overrepresentation of White faces enabled the model to generate faces that represented the average of all those features in a way that made them look more human than human.
The paper concluded that this bias in the training datasets raises important issues. “If AI faces do appear more realistic for White faces than other groups, their use will confound perceptions of race with perceptions of being ‘human’”, it noted.
If an AI model’s idea of a human face is an unnatural average of White facial features, how would it distinguish between a real human of different ethnicity and an AI fake if asked to do so?
Senior author Dr. Amy Dawel said, “If white AI faces are consistently perceived as more realistic, this technology could have serious implications for people of color by ultimately reinforcing racial biases online.”
What features led to the mistakes?
There must be some specific features of an AI face that make us think it’s more human than a real human face. In the second experiment, 610 participants were asked to rate the AI and human faces on a range of 14 attributes including features like attractiveness, eye contact, and expressiveness.
Combining this data with that of the first experiment enabled the researchers to identify what made people more likely to identify a face as being AI-generated or real.
They found that the hyperrealism of AI faces could be attributed to them being “significantly more average (less distinctive), familiar, and attractive, and less memorable than human faces.”
The fact that we’re so quick to accept that an AI face is real shows how important it is to have AI fake detection tools.
The researchers took the data from the human-perceived attributes and how they were used correctly and incorrectly when misidentifying AI faces and created a machine learning model to spot AI faces. The model was able to accurately classify face type with 94% accuracy.
We’re unlikely to run a face through an AI face checker each time we see one online. And face generators are only going to get better at beating the fake detectors.
Dr. Dawel summed up what our best option is in the face of this: “Educating people about the perceived realism of AI faces could help make the public appropriately skeptical about the images they’re seeing online.”
If we remind ourselves that we’re really bad at spotting fakes, maybe we’ll be less likely to be fooled by them.
