Images of faces generated by artificial intelligence (AI) are so realistic that even “super recognizers” — an elite group with exceptionally strong facial processing abilities — are no better than chance at detecting fake faces.
People with typical recognition capabilities are worse than chance: more often than not, they think AI-generated faces are real.
“I think it was encouraging that our kind of quite short training procedure increased performance in both groups quite a lot,” lead study author Katie Gray, an associate professor in psychology at the University of Reading in the U.K., told Live Science.
Surprisingly, the training increased accuracy by similar amounts in super recognizers and typical recognizers, Gray said. Because super recognizers are better at spotting fake faces at baseline, this suggests that they are relying on another set of clues, not simply rendering errors, to identify fake faces.
Gray hopes that scientists will be able to harness super recognizers’ enhanced detection skills to better spot AI-generated images in the future.
“To best detect synthetic faces, it may be possible to use AI detection algorithms with a human-in-the-loop approach — where that human is a trained SR [super recognizer],” the authors wrote in the study.
Detecting deepfakes
In recent years, there has been an onslaught of AI-generated images online. Deepfake faces are created using a two-stage AI algorithm called generative adversarial networks. First, a fake image is generated based on real-world images, and the resulting image is then scrutinized by a discriminator that determines whether it is real or fake. With iteration, the fake images become realistic enough to get past the discriminator.
These algorithms have now improved to such an extent that individuals are often duped into thinking fake faces are more “real” than real faces — a phenomenon known as “hyperrealism.”
As a result, researchers are now trying to design training regiments that can improve individuals’ abilities to detect AI faces. These trainings point out common rendering errors in AI-generated faces, such as the face having a middle tooth, an odd-looking hairline or unnatural-looking skin texture. They also highlight that fake faces tend to be more proportional than real ones.
In theory, so-called super recognizers should be better at spotting fakes than the average person. These super recognizers are individuals who excel in facial perception and recognition tasks, in which they might be shown two photographs of unfamiliar individuals and asked to identify if they are the same person or not. But to date, few studies have examined super recognizers’ abilities to detect fake faces, and whether training can improve their performance.
To fill this gap, Gray and her team ran a series of online experiments comparing the performance of a group of super recognizers to typical recognizers. The super recognizers were recruited from the Greenwich Face and Voice Recognition Laboratory volunteer database; they had performed in the top 2% of individuals in tasks where they were shown unfamiliar faces and had to remember them.
In the first experiment, an image of a face appeared onscreen and was either real or computer-generated. Participants had 10 seconds to decide if the face was real or not. Super recognizers performed no better than if they had randomly guessed, spotting only 41% of AI faces. Typical recognizers correctly identified only about 30% of fakes.
Each cohort also differed in how often they thought real faces were fake. This occurred in 39% of cases for super recognizers and in around 46% for typical recognizers.
The next experiment was identical, but included a new set of participants who received a five-minute training session in which they were shown examples of errors in AI-generated faces. They were then tested on 10 faces and provided with real-time feedback on their accuracy at detecting fakes. The final stage of the training involved a recap of rendering errors to look out for. The participants then repeated the original task from the first experiment.
Training greatly improved detection accuracy, with super recognizers spotting 64% of fake faces and typical recognizers noticing 51%. The rate that each group inaccurately called real faces fake was about the same as the first experiment, with super recognizers and typical recognizers rating real faces as “not real” in 37% and 49% of cases, respectively.
Trained participants tended to take longer to scrutinize the images than the untrained participants had — typical recognizers slowed by about 1.9 seconds and super recognizers did by 1.2 seconds. Gray said this is a key message to anyone who is trying to determine if a face they see is real or fake: slow down and really inspect the features.
It is worth noting, however, that the test was conducted immediately after participants completed the training, so it is unclear how long the effect lasts.
“The training cannot be considered a lasting, effective intervention, since it was not re-tested,” Meike Ramon, a professor of applied data science and expert in face processing at the Bern University of Applied Sciences in Switzerland, wrote in a review of the study conducted before it went to print.
And since separate participants were used in the two experiments, we cannot be sure how much training improves an individual’s detection skills, Ramon added. That would require testing the same set of people twice, before and after training.
