Spotting AI deepfakes a ‘cat and mouse game,’ top expert warns

• This article contains various photographs of people. While reading, try to guess which ones are real and which ones are created by artificial intelligence.

The expert behind a new training program designed to help people debunk artificial intelligence deepfakes has admitted that we find ourselves in a “cat-and-mouse game” with advancing technology.

The Australian National University has launched a new scheme that it says makes people 40 per cent better at extracting AI-generated faces from human photographs.

But the lead researcher behind the program admitted it’s only a matter of time before technology finds new ways to avoid getting caught.

The state-of-the-art program begins by giving participants an ID card consisting of mugshots, some artificial intelligence, some real, and tasks them with separating the real from the fake.

Participants are then given a short course on how to spot differences and are asked to repeat the task at the end.

The end result was about a 40 percent improvement in people’s recognition skills, said lead researcher Amy Dawel.

“We found that even people who were super recognizers were not as successful as they initially expected,” Professor Dawel said. “We often find that people are surprised at how terrible they are at it and how much they’ve improved.

“The more confident they are, the more mistakes they tend to make.”

Professor Dawel said that as technology progresses, the task becomes increasingly difficult and it becomes increasingly difficult to detect images created by artificial intelligence.

Gone are the days of simple markings such as extra fingers on hands or mismatched earrings, he said.

“AI-generated faces often look pretty average and don’t stand out from the crowd, which can make them more attractive,” he said.

“This means that real humans have become really terrible at detecting AI genes.”

The research project behind the tutorial revealed six keys to rooting out fakes: distinctiveness, memorability, proportionality, symmetry, attractiveness and expressiveness.

Professor Dawel said symmetry, in particular, was a strong sign that the image was artificial intelligence.

“AI-generated faces tend to be much more symmetrical, while real people tend to be more interesting and have individual features,” he added.

cat and mouse game

Professor Dawel said it was becoming increasingly important for people to root out AI deepfakes because of the ways they could take advantage of the technology in the dark corners of the internet.

He said AI fraud is estimated to cost the global economy US$40 billion by the end of next year.

“There are many ways AI-generated faces can be used and abused, whether that’s catfishing dating sites, accessing bank accounts or forging passports,” he said.

“People are now able to hide behind this and claim that things that are actually real are fake.

“Being able to tell the difference is very important.”

But the associate professor also admitted that as AI technology gets better at detecting deepfakes, it will inevitably get better at avoiding detection.

“I definitely think it’s a cat-and-mouse game and there’s no escaping it,” he added.

“But in the same way that AI itself is getting better, so are AI detectors, but human psychology is also an important thing to rely on.”

‘It’s getting harder and harder’

Professor Toby Walsh from the University of NSW said rapidly advancing technology was regularly forcing people to question what was in front of them.

“The real challenge is that we can no longer believe our own eyes,” he said. “I think we’ll continue to see things that aren’t real, but we’ll also continue to question things that are real.

“The way to combat this is through education but also through digital watermarks.”

Digital watermarks are invisible, machine-readable identifiers attached to files such as photos that reveal their origin.

Professor Walsh said as technology continues to advance these could potentially be the only way to distinguish real from deepfake.

“It’s getting really difficult to distinguish between real and fake, technology is advancing so quickly,” he said.

“Another problem is that it gives people a way to avoid accountability.

“There are politicians who say terrible things and then dismiss them as deepfakes; that’s all their supporters need to hear to believe them.”

‘Arms race’

Wolfgang Mayer, associate professor of computer science at the University of Adelaide, said progress in artificial intelligence technology was achieved thanks to competition between the world’s largest technology companies.

“It’s come a long way in the last few years, and I think that’s driven by an arms race between the big tech companies that are determined to be the best and produce the most content,” he said.

“Obviously they don’t set out to create deepfakes, but this is an obvious byproduct of existing technology.

“The technology is available and easy to use, for better or worse.”

He added that there is a greater need for people to learn to question the sources of things they see online.

“Teaching people how to spot deepfakes is a useful thing in the short term, but in the long term it’s much more important to raise awareness that not everything you see on the internet is real,” Professor Mayer said.

He added that technology has advanced to the point where rooting it out is about looking at what isn’t there rather than what is there.

“Deepfakes are almost becoming too perfect,” he said.

“We used to be able to easily distinguish things like eyes, mouths or teeth, for example. Now you have to look for flaws that are barely there, flaws that exist in real people.”

What does Artificial Intelligence say?

As university experts share tips for spotting AI-generated faces, will technology be ready to show its hand, too?

NewsWire turned to Google’s Gemini AI chatbot to ask for some clues, which provided a variety of pointers.

• hands and feet – Look for extra fingers, missing joints, unnaturally long fingers, or hands that turn into grasped objects

• ‘Plastic’ or airbrushed skin – look for natural imperfections such as prints, fine lines, moles and light blemishes to find the real person. AI images can often look like heavy beauty filters

• ‘Spaghetti’ hair – Look closely at where the hairline meets the forehead. Artificial intelligence often struggles with realistic hair follicles

The chatbot added that the most effective way to spot a fake is to look closely. “Always zoom in when trying to spot an AI image,” he said. “AI images look incredibly convincing as thumbnails on a smartphone screen, but zooming in 200 percent to 300 percent on eyes, ears, and hands almost always reveals digital seams.” To express your interest in training, visit: https://tinyurl.com/ai-face-study-register • Photo 3 is by Professor Dawel, photo 5 is by Professor Mayer and photo 7 is by Professor Walsh. The rest were all created using StyleGAN3 and are not real people.