At first glance, these AI-generated faces look real. But if you look closer, their features are a little too perfect and too symmetrical.
Artificial faces are harder than ever to identify with the human eye thanks to more sophisticated artificial intelligence models.
But a study from the Australian National University has shown humans can be trained on how to sort out real faces from AI-generated deepfakes with remarkable accuracy.
Associate Professor Amy Dawel said their training approach improved the participants' accuracy in detecting AI faces from 40 per cent to 80 per cent.
"This study were surprising to us, that it improved their performance so much," Dr Dawel said.
"The fact that we got such a massive boost was just astounding to us."
In the early days of AI-generated images, there were tell-tale signs of a fake.
"You go back five or six years ago and it was having a sixth finger or something weird and strange going on," Dr Dawel said.
"But over time, what's happened is that these faces have become more and more realistic, so that the obvious signs aren't there."
The study participants were brought to the Emotions and Faces Lab at ANU where they were tested on their abilities before and after training from ANU honours student Tanya George.
The participants were exposed to about 100 different faces, half human and half AI-generated, and were asked to rate them on the six traits that distinguish AI images. The study used StyleGAN faces, which are some of the most convincing fakes available.
"These faces tend to be a bit more symmetrical, they tend to be a little bit less expressive, they tend to be more in proportion," Dr Dawel said.
"The problem is that we think that things that look odd are likely to be AI, but in fact humans often have quite unusual, distinct features, and so we're flipping what we're thinking there."
After being trained the participants were tested again. They were shown three images and had to select the one that was the AI-generated face. The best-performing participants got close to 100 per cent accuracy in detecting the AI deepfakes.
Slide to reveal which faces are real and which are AI-generated. These images were used in a previous study by Sophie Nightingale and Hany Farid.
Some argue that algorithms, not humans, are best-placed to detect an AI fake. But humans don't always know what the algorithms are doing and they may only work on certain types of faces.
Dr Dawel said algorithms and human skills should be paired together to ensure there is an accountable way to make decisions.
"We're living in an age where AI is becoming part of our everyday online experience and so being able to understand that we're actually, first of all, very poor at doing this is important for us in terms of making sure we're double-checking our facts.
"But second of all, in high-stakes situations if we can train humans to be able to do this really effectively, it gives us an explainable and accountable way of understanding and keeping humans in the loop."
The findings of the study were replicated by a team led by Professor Jim Tanaka and Dr Eric Mah at the University of Victoria, Canada.
The study is published in the scientific journal PNAS.
The lab is continuing its research to look at whether the training could be used for other types of media, such as voice recordings and video, and whether it is effective for vulnerable groups, such as children and older adults.
The challenge for regular people is that the AI image generators continue to evolve and get more sophisticated. But so far, the platforms have been unable to replicate the imperfect nature of the human face.
"It's tricky for them to do that in a way that doesn't make the images more bizarre overall," Dr Dawel said.
"There's some things that are inherent in these algorithms that mean that they're producing the most common patterns in their training data, which is part of what makes them a little bit more symmetrical and more proportional. They tend towards the more average faces, so we call it being hyper average."
