I Can’t Hide a Lie: A.I.
AI-generated content is rapidly permeating the internet, and it’s becoming hard to distinguish from reality. A recent study published in a scientific journal showed that, without specialized training, most people struggle to tell apart human faces from those created by AI. Researchers noted that generative adversarial networks (GANs) can produce highly realistic synthetic faces, which could be misused.
For instance, TikTok users recently called out a deepfake doctor for misleading them with baseless medical claims.
“It’s pretty gratifying that a short training session like ours yielded a notable enhancement in the performance of both groups,” said Katie Gray, the study’s lead author and an associate professor of psychology at the University of Reading in the UK.
These AI-generated faces can be so convincing that some people may even find them to appear more genuine than actual human faces. To combat this, researchers are developing a five-minute training guide aimed at helping individuals recognize AI-generated imposters.
The training focuses on spotting flaws in the artificial faces, like missing teeth, unusual hairlines, or unnatural skin textures. Interestingly, these synthetic faces are often more symmetrical than real ones.
In their experiments, the researchers assessed the facial recognition abilities of ordinary individuals compared to “super-recognizers,” who excel in identifying faces. Participants were selected from a volunteer database and had shown impressively high scores in remembering new faces.
In the initial test, participants were shown a face for 10 seconds and asked to determine if it was real or fake. The standard recognizers could only identify 30% of the fakes, whereas the super recognizers managed just 41%, which is still below what pure guessing would achieve.
The second experiment followed a similar format, although this time, a new group underwent the five-minute training to help them spot errors in AI-generated faces. Each participant evaluated ten faces, with real-time accuracy measures taken and errors discussed afterward.
In this follow-up test, the super-recognizers identified 64% of the fake faces, while the average individuals improved their detection rate to 51%—a meaningful enhancement. Those who were trained also seemed to take their time in analyzing faces before responding.
“I find it reassuring that with our brief training session, both groups showed considerable improvement,” Gray commented.
However, there are some limitations to consider. The participants were assessed right after training, which raises questions about how well they would perform without a refresher later on.
Still, with so many AI impersonators surfacing online, it’s critical to equip people with the skills necessary to differentiate between humans and AI.
Moreover, the capabilities of this technology extend beyond simple visuals. Research has indicated that language models have become indistinguishable from human speakers, essentially passing the Turing test.
