Advertising
You may have seen photos that suggest otherwise, but former President Donald Trump wasn’t arrested last week, and the pope wasn’t wearing an elegant crisp white puffer coat. These recent viral hits have been the fruit of artificial intelligence systems that process a user’s text prompt to create images. They demonstrate how these programs have become very good very quickly – and are now convincing enough to fool an unwitting observer.
So how can skeptical viewers recognize images that may have been generated by an artificial intelligence system such as DALL-E, Midjourney, or Stable Diffusion? Each AI image generator—and each image from a given generator—differs in how convincing it can be and what telltale signs might give its algorithm away. For example, AI systems have historically had difficulty mimicking human hands and have produced garbled appendages with too many digits. However, as technology improves, systems like Midjourney V5 seem to have cracked the problem – at least in some examples. In general, experts say that the best images from the best generators are difficult, if not impossible, to distinguish from real images.
“It’s pretty amazing what AI image generators can do,” says S. Shyam Sundar, a researcher at Pennsylvania State University who studies the psychological effects of media technologies. “In terms of image generation capabilities, there’s been a huge leap in the last year or so.”
Advertising
Some of the factors behind this leap in capability are the ever-growing number of images available to train such AI systems, as well as advances in computing infrastructure and interfaces that make the technology accessible to ordinary internet users, Sundar notes. The result is that man-made images are everywhere and can be “almost impossible to see,” he says.
A recent experiment showed how well AI can deceive. Sophie Nightingale, a psychologist at Lancaster University in England who focuses on digital technology, co-authored a study that tested whether online volunteers chose between passport-like headshots created by an AI system called StyleGAN2 and real ones images could distinguish. The results were discouraging, even in late 2021 when researchers conducted the experiment. “On average, people were pretty random,” says Nightingale. “Basically, we’ve gotten to a point where it’s so realistic that people can’t reliably tell the difference between these synthetic faces and real, real faces — faces of real people who really exist.” AI provided some help (researchers sorted the images generated by StyleGAN2 to select only the most realistic), Nightingale says that someone wanting to use such a program for nefarious purposes would probably do the same.
In a second test, the researchers tried to help the test subjects improve their AI recognition skills. They marked each answer true or false after the participants answered, and they also prepared the participants in advance by having them read through advice on recognizing artificially generated images. This advice highlighted areas where AI algorithms often stumble, creating things like mismatched earrings or blurring a person’s teeth. Nightingale also notes that algorithms often struggle to create anything more sophisticated than a simple background. But even with those additions, participants’ accuracy only increased by about 10 percent, she says — and the AI system that generated the images used in the study has since been upgraded to a new and improved version.
Advertising
Ironically, as image generation technology continues to improve, human’s best defense against being fooled by an AI system might be another AI system: one trained to recognize artificial images. Experts say that as AI imaging advances, algorithms are better equipped than humans to recognize some of the tiny, pixel-sized fingerprints of robot creation.
Building these AI detective programs works just like any other machine learning task, says Yong Jae Lee, a computer scientist at the University of Wisconsin-Madison. “You’re collecting a dataset of real images, and you’re also collecting a dataset of AI-generated images,” says Lee. “Then you can train a machine learning model to distinguish the two.”
Still, these systems have significant flaws, say Lee and other experts. Most of these algorithms are trained on images from a specific AI generator and are unable to identify fakes generated by other algorithms. (Lee says he and a research team are working on a way to avoid this problem by training the detector to recognize what makes up an image instead real.) Most detectors also lack the user-friendly interfaces that have enticed so many people to try out the generative AI systems.
Advertising
Additionally, AI detectors will always try to keep up with AI image generators, some of which contain similar detection algorithms but use them to learn how to make their fake output less detectable. “The battle between AI systems that generate images and AI systems that recognize the images generated by AI is becoming an arms race,” said Wael AbdAlmageed, associate professor of computer science at the University of Southern California. “I don’t see any team winning in the foreseeable future.”
AbdAlmageed says no approach will ever be able to capture every single artificially generated image – but that doesn’t mean we should give up. He suggests that social media platforms need to start confronting AI-generated content on their websites, as these companies are better able to implement detection algorithms than individual users.
And users need to be more skeptical about visual information, asking if it’s fake, AI-generated, or harmful before sharing it. “We as a human species kind of grow up thinking that seeing is believing,” says AbdAlmageed. “That’s not true anymore. Seeing is no longer believing.”
