AI image generators have made huge strides in creating realistic faces, often looking more authentic than actual people. A new study from researchers in the UK shows there’s a way to improve how we spot these AI-generated images.
The researchers tested 664 volunteers, including “super-recognizers” who excel at identifying faces and those with average face-recognition skills. Both groups found it challenging to recognize AI faces, though the super-recognizers performed better. Interestingly, after just five minutes of training, the super-recognizers improved even more.
Eilidh Noyes, a psychology researcher at the University of Leeds, explains, “AI images are easier to create and harder to detect. They can be misused, making it essential to find ways to identify them.”
In the study, participants faced two tasks: identifying a single face as real or AI and comparing a real face with an AI one. Those who didn’t receive training correctly identified AI faces only 41% of the time, while the average participants only managed 31%. Given half the images were AI-generated, this means many were fooled, highlighting how realistic these images can be.
When training was provided, average recognizers achieved an accuracy of 51%, while super-recognizers reached 64%. They learned to look for key details like missing teeth or odd blurring. Noyes shares that super-recognizers, when trained, can assist in spotting AI faces better.
AI typically uses a process called a generative adversarial network (GAN). This involves two algorithms: one creates faces, while the other checks them against real human features, driving the generator to produce lifelike results.
AI images are popping up everywhere—from fake dating profiles to scams. Simple training can empower more people to spot misleading images. Katie Gray, another psychology researcher, notes that this brief training can make a real difference in verifying identities online.
This study’s findings could be crucial in our tech-driven world where being able to discern real from fake is increasingly important. For more on this research, check out the details in the Royal Society Open Science.
Source link
MSFT Content
