On social networks, perfect faces parade, smile, pose as if they came out of a family album. Some belong to real people, others never existed. Many swear that they would no longer know
recognize a face created by an AI. Is it really lost for everyone?
A team from the Vanderbilt University has just published a study that turns this idea on its head. Using a new test, theAI Face Testresearchers show that the key is neither your IQ nor your computer skills, but a more discrete visual ability: the object recognition. And there, the differences between individuals become striking.
Recognizing a face created by an AI, a very unevenly distributed talent
The authors, including psychologist Isabel Gauthier, were interested in our way of distinguishing a real face from one generated by AI, in a context of deepfakes and fake news visual. Isabel Gauthier summarizes: “I think a lot of messaging suggests that we can’t tell the difference, when in reality it’s a distribution of people. Some people don’t see the difference, others do very well, and still others do quite well. As AI becomes more and more a part of our reality, I think it’s helpful to know that some people are better at it than others“, estimates Isabel Gauthier, cited by Vanderbilt University. In reality, performances range from very poor… to excellent.
To understand this variation, the team used a concept already studied in psychology: a general ability to recognize objects, that is to say, to distinguish visually similar objects. It is involved in very different tasks, such as the detection of pulmonary nodules on x-rays, the identification of cancer cells, the recognition of musical scores, etc. The researchers point out that “humans with stronger object recognition skills consistently outperform others at identifying AI-generated faces, and their performance remains stable when retested“, note the authors of the study published by Vanderbilt University.
AI Face Test: the test that measures your object recognition
Concretely, the AI Face Test presents a long series of faces, some very real, others produced by image models. Each time, you have to decide: true or AI. “We didn’t just want to examine whether people are able to differentiate between a real face and an AI-generated face, but to compare people on their ability to perform this task and see if we could predict their performance using object recognition.“, specifies Isabel Gauthier. This overall capacity turns out to be the best predictor of scores.
Another striking result: neither general intelligence, nor experience with AI, nor even specific face recognition skills are enough to explain who is doing well. “These results highlight a visual ability that has very general applications“, explains Isabel Gauthier. “This is a stable characteristic that helps individuals address new perceptual challenges, including those created by AI. We were surprised to find that neither intelligence nor even technological training could accurately determine whether a face was generated by AI“.
Why this visual “superpower” matters in the face of misinformation
The authors also note that understanding these individual differences can be used to better organize interactions between humans and generative AI systems, for example to select people who perform well in validating sensitive images or to refine the datasets used to train models. A way of reminding us that, faced with faces created by AI that are ever more credible, our human gaze retains a strategic place… even if we do not all have the same talent for distinguishing truth from falsehood.