At first glance, the two photographs looked completely convincing. A guitarist posed with his instrument, a tennis player mid-swing with racket in hand—nothing obviously wrong. But to Groh, an assistant professor of management and organizations at Kellogg, the images were clearly fake. The guitar’s fret inlays sat in unusual positions, and the tennis racket strings curved in ways no real racket would ever allow. These small, almost invisible details are the kind of tells that Groh has learned to spot through years of studying AI-generated images. Most people, though, don’t have that kind of trained eye. And in these early, fast-moving days of generative AI, it’s still unclear whether that skill can be taught quickly—or whether we are all doomed to be fooled by increasingly realistic synthetic pictures. Groh wanted to find out. So he and his colleagues set up a simple but important experiment: could a short, focused training session help people get better at telling real photos from AI-generated ones? And not just any people. They focused on professionals whose jobs depend on making these judgments in some of the highest-stakes situations imaginable: intelligence analysts working for the U.S. government. These are the people who sift through images every day, trying to figure out what is real and what is a lie, often with national security hanging in the balance. If a fake image could trigger a drone strike, a false accusation, or a diplomatic crisis, the cost of being fooled is enormous. Groh gave these analysts a single half-hour training session, pointing out common patterns in real versus AI-generated images, and then tested them again. The results were reassuring. The training worked. It significantly improved the analysts’ overall accuracy, proving that even in the age of hyperrealistic fakes, we are not helpless. “It’s possible to get better,” Groh says. “We’re not doomed not to be able to tell real from fake.”
The latest AI models have made it possible for nearly anyone to create a realistic-looking picture simply by typing a text prompt. You don’t need artistic talent, expensive equipment, or even much technical know-how. You just describe what you want, and the AI generates it. In just a few years, these models have improved so dramatically that the conventional wisdom has become: people simply cannot tell the difference anymore. But Groh’s earlier research suggested that wasn’t entirely true. In a study of ordinary people, he and his colleagues found that everyday participants could correctly distinguish real from fake images about three-quarters of the time. They weren’t perfect—far from it—but they were also nowhere near random guessing. And an interesting pattern emerged: the longer people looked at an image, the more likely they were to spot the fake. Careful observation helped. But even with that natural ability, there was still plenty of room for improvement, especially for professionals like intelligence analysts who deal with sensitive, high-stakes material every day. If an analyst misjudges an image, the consequences could be severe. Imagine a fake photo suggesting a terrorist has been spotted in a particular location. If officials fail to recognize that the image is AI-generated, they might authorize a drone strike without just cause. That kind of mistake is catastrophic. So Groh and his colleagues designed a 30-minute training session aimed at closing the gap. The training walked participants through patterns and tell-tale signs found in 7 real and 50 AI-generated images. It taught them what to look for, and what to be suspicious of. For example, AI images often look overly perfect, almost too cinematic to be true. They feature people with symmetrical, classically beautiful features, flawless skin, and lighting that feels staged. Other common signs are more specific: waxy or glossy skin, missing teeth, or impossible little details like a doctor’s stethoscope merging into a single loop instead of two separate earpieces. These are the kinds of clues that can give a fake away—once you know to look for them.
For the study, Groh and his colleagues worked with 32 intelligence analysts from various U.S. agencies—people who write daily briefs for the president and other senior White House officials. These are not casual social media users; they are professionals whose entire job revolves around interpreting visual information and making decisions that affect national security. The research team asked the analysts to look at 40 images on a customized web interface and classify each one as either real or AI-generated. The analysts also had to write down the reason behind each decision. The image set included a variety of types: portraits, full-body shots, posed group photos, and candid snapshots. Importantly, the images were paired by content. For example, the set might include a real photograph of a female astronaut and an AI-generated image of a female astronaut. But the participants were shown each image one at a time, so they couldn’t simply compare two versions side by side. They had to rely on their instincts, their knowledge, and whatever clues they could spot. As a group, the analysts correctly identified 73 percent of the images before the training. That’s almost exactly the same level of accuracy that Groh’s team had measured earlier in ordinary people. That was a striking finding. Even the people who professionally deal with sensitive images every day were no better than the average person at spotting AI-generated fakes. The AI models are so new, and so good, that the expertise analysts have developed for recognizing traditional photo manipulation or staged images doesn’t automatically transfer to this new challenge. But then came the training. Groh presented a 30-minute slide deck designed to teach the analysts how to distinguish real from AI-generated images. After the session, the analysts repeated the same online classification exercise with a different set of 40 images. The improvement was clear and meaningful. Before training, their accuracy hovered around 73 percent. After training, it jumped to 82 percent—a nine-point gain. Even more impressive, their ability to correctly classify both images in a pair, such as the real and fake versions of an astronaut, improved by 11 percent. In a world where fakes are everywhere and getting harder to spot, that kind of improvement could be the difference between catching a lie and being caught by it.
One of the most interesting findings was what drove the overall improvement. The gain was mostly driven by a 14 percent increase in accuracy at identifying real images as real. In other words, the training helped the analysts reduce their false positives—the tendency to incorrectly label authentic photographs as AI-generated. This may not sound as dramatic as catching fakes, but it is actually the hardest and most important part of the battle. With an AI-generated image, you can be confident it’s fake once you spot a clear artifact, like a warped hand or a nonsensical background. But a real image doesn’t have those obvious tells. It doesn’t have a specific artifact to point to; it simply has the absence of artifacts. That makes it much harder to confirm that something is genuine. If people are too suspicious, they start assuming everything is fake, which is its own kind of danger. Groh notes that in some previous training experiments by other researchers, participants did get better at identifying AI images, but they also became more likely to accuse real photographs of being AI-generated. That’s a serious problem. It creates a culture of distrust where no visual evidence can be trusted, which is just as damaging as being fooled by a fake. “That’s kind of problematic,” Groh says. “Now you’re just leading everyone to think everything’s fake.” The training in this study avoided that trap, and the reason seems to be that it gave analysts a deeper understanding of what real images actually look like—not just what fakes look like. The evidence appeared in their written justifications. The second time they took the test, the analysts gave richer, more detailed reasons for their choices. Before training, their comments were vague and uncertain. They might write “hair” or “fingers” or simply say “something looks off.” After training, they were much more specific, pointing out details that reflected key patterns: a door that opened into empty space, dirt that looked too uniform, lighting that didn’t match the scene. They weren’t just guessing or reacting to a gut feeling anymore. They knew what to look for, and they knew why it mattered.
Beyond national security, the ability to tell real images from AI-generated ones matters in almost every area of life. It matters for spotting fake news, for verifying evidence, for making sound business decisions, and for maintaining basic trust in the world around us. Trust, as Groh points out, is fundamental to business. Consider a budding entrepreneur pitching a product to investors, showing photos of a prototype or a team or a manufacturing facility. If investors can’t tell whether those images are real, they can’t make informed decisions. If a customer can’t tell whether a product photo is genuine, they may buy something that doesn’t exist. “The moment that you can’t tell the difference at all is the moment you can no longer trust any visual medium,” Groh says. That is a frightening prospect. In an ideal world, policymakers would step in and create guardrails around AI image generation—requiring companies to clearly disclose when images are fake, for example, or making it illegal to create certain kinds of deceptive content. But that kind of policymaking hasn’t caught up with the technology yet. It’s not here yet, and Groh doesn’t expect it to arrive overnight. In the meantime, individuals need to build this skill themselves. The good news is that it’s possible. The training in this study was short—just 30 minutes—and it produced significant, meaningful improvement. But there are other ways to get better too. One of the best methods is simply to experiment with AI models personally. The more you use these tools, the more familiar you become with their tendencies, their quirks, and their limitations. You start to notice that AI has a certain look, a certain way of organizing details that real life rarely replicates. You start to see the patterns that give fakes away. As Groh puts it, “The biggest thing is playing around with these tools themselves. You start to see the limitations.”
So, are we doomed to drown in a sea of indistinguishable fakes? Not necessarily. The research from Groh and his colleagues offers a genuinely hopeful message. Human beings are not passive victims of technology. With a little guidance, we can improve our judgment, sharpen our instincts, and learn to spot the cracks in the machine. The intelligence analysts in this study went from being no better than the general public to being significantly more accurate after just half an hour of training. That improvement didn’t come from superhuman abilities or years of specialized study. It came from knowing what to look for, and from understanding that even the most sophisticated AI still leaves traces of its artificiality. But there’s a cautionary note hidden in the findings as well. The skill is learnable, but it isn’t automatic. It requires effort, curiosity, and a willingness to engage with the technology instead of ignoring it. It also requires balance. We can’t become so paranoid that we reject everything as fake, nor can we be so naive that we accept everything as real. The goal is not to become perfect lie detectors. The goal is to get better, to reduce the margin of error, and to make it harder for malicious actors to exploit our ignorance. Groh’s work suggests that this is possible. We are not doomed. But we are responsible. In an era where anyone can generate a convincing fake image in seconds, the ability to tell real from fake is no longer just a nice skill to have—it’s a survival skill. And the good news is that it’s one we can all develop, one training session at a time.