Meta AI Image Detector Misses Some of Its Own Cropped AI-Generated Images

When it comes to discerning artificial intelligence from human-generated media, Meta appears to be tripping itself up-and, in a more worrisome way, tripping up its own tech. The tech titan’s image detection system is demonstrably flawed, failing to identify a portion of the AI-generated images its own algorithms produced after the images have been cropped, according to researchers who conducted tests on the company’s technology. The failure points to an alarming limitation that continues to plague technology that’s designed to weed out synthetically generated content.

Even as AI image generators are getting smarter-and indistinguishable from real photographs to the human eye-the technology tasked with distinguishing between real images and manufactured ones lags behind, at least in some cases, the research from a group of academics suggests.

Meta has been investing heavily in technologies that enable detection of AI-generated imagery. It’s crucial that we are transparent with users about when content has been generated by artificial intelligence, Meta said in the press release it posted in September when announcing that it expanded its AI detection capabilities, a year-long process that included new technology that has shown an ability to detect 90 percent of the synthetic images that the company's AI models produced. But those new detectors are not foolproof, the new academic research claims, showing what it calls a gap that Meta and other tech companies and research organizations will have to address. Some AI-generated images can defeat AI-based image detectors when they've been "re-processed," a common scenario when users crop and share images online-something for which the AI tools do not appear to be adequately designed to protect.

What the experts discovered Meta's latest AI detectors are also hampered when dealing with "unqualified training examples," in the words of a paper that has yet to be peer-reviewed.

Those are images that an AI training dataset may be unable to consistently label-often because they’ve undergone a process to make them more realistic, thereby skewing how the model was trained to identify a synthetic image. Those types of errors can be more significant, said one of the academics behind the new study, though Meta's study said its detectors still flagged most synthetic examples produced for the test and even those not produced by its algorithms as AI images. "If such examples exist in your training data and if your method is not robust enough to these types of corruptions then you may miss some AI-generated images," said Iryna Boyadjis, an assistant research scientist at Rice University.

While Meta may be overstating the real-world impact of these errors, such instances highlight a challenge for tech companies attempting to keep pace with advancements in AI imagery. No solution to detection is likely to last forever, they explain. For now, researchers advise, users can look for subtle visual artifacts in images that may reveal their artificial nature-something that will likely become increasingly difficult, and eventually impossible, as the AI technology advances.

For example, when you are consuming a piece of information online, consider not only the words in it but the image it depicts.

If it’s one of the few anomalies - where an AI-generated image has made its way through, without a clear identifying label - some of the inconsistencies can hint at their artificial origins. But for users to rely solely on themselves to identify potential deep fakes may be the only option at present, with a recent study of large image models indicating their vulnerability to subtle but simple editing like changing an image file’s properties, or even cropping it. Meta already implements visual watermarks, invisible patterns added to the pixels of a synthetic image to help distinguish it from real ones, in some of its content creation tools, and the company has expanded its AI-generated detection model this month with new methods that “allow us to better detect when content has been generated by the same AI system used to create them,” its head of responsible AI AI technology, Lena Petrova, wrote on the company's platform on X, formerly known as Twitter.