Microsoft has launched a new online game, “Real or Not,” challenging players to differentiate between AI-generated visuals and authentic snapshots. The game, developed in collaboration with Cornell University researchers, draws from a study that reveals how difficult it can be for humans—and even machines—to spot the difference. Try the game here: https://www.realornotquiz.com/.
The study, conducted with 12,500 participants worldwide who evaluated 287,000 images, found that people correctly identified AI-generated images only 63 percent of the time. Intriguingly, the images most likely to be mistaken for fakes were genuine photographs, particularly those depicting U.S. military scenes in unusual lighting or settings. These conditions, the researchers noted, often confused participants, highlighting the limitations of human perception in an age of sophisticated AI.
Microsoft’s game invites users to test their own abilities, offering a series of images and asking players to determine whether they are real or AI-crafted. The results can be humbling: Many players, including seasoned tech observers, score no better than chance. In contrast, Microsoft claims its AI detection tools achieve a 95 percent accuracy rate in identifying synthetic images, though independent verification of this figure remains pending.
The Cornell study sheds light on why some AI-generated images are so convincing. Images created using generative adversarial networks (GANs), particularly those depicting human profiles, proved the most deceptive, with only 65 percent of participants correctly identifying them as fakes. Landscapes and cityscapes, often generated with generic filenames to mimic authentic photos, were correctly identified as AI creations in just 21 to 23 percent of cases. These images matched real photographs in noise levels, brightness, and entropy—a measure of visual randomness—making them nearly indistinguishable to the untrained eye.
Thomas Roca, the study’s lead author, emphasized the broader implications. “The findings underscore the urgent need for watermarking and robust detection tools to combat AI-generated misinformation,” he said. Without such measures, the spread of convincing deepfakes could erode trust in visual media, a concern amplified by the study’s warning that understanding what fools humans could make future disinformation campaigns even more effective.
Other platforms, like Wasitai.com, claim to detect AI-generated content, but reports suggest many overstate their accuracy. As AI image generation advances, the challenge of separating fact from fiction grows more pressing. Microsoft’s “Real or Not” game, while an engaging experiment, serves as a reminder of the stakes involved in a world where seeing is no longer believing.
