Instagram’s system labels real photos as AI content

    Instagram’s system for detecting AI-generated content has recently begun incorrectly adding the “AI Content” label to original photos. According to users, images that were neither generated nor edited using AI are also being labeled. At the same time, some images created entirely with AI remain unlabeled. This raises questions about the label’s reliability.    Some of the complaints involve Canva’s background removal tool. In some cases, removing a background or cleaning up a small blemish in a photo caused the entire image to be labeled as AI content. Although these tools use machine-learning technology, they differ from systems that create entirely new images from text. For some photos, it is unclear why the label was added.    A similar problem occurred in 2024. Meta said at the time that it checked IPTC and C2PA metadata. This data consists of technical records about a file’s origin and the changes made to it. Photos with minor edits made using AI tools in Adobe software were also given the “Made by AI” label. The company later said it would adjust its labeling based on the extent to which AI had been used in an image.    Canva acknowledged that some assistive AI tools had mistakenly been classified as generative systems, meaning systems that create new content. The company said the problem had been resolved. The platform’s help page also states that the background removal feature does not add metadata to a design indicating that it was created with AI. Nevertheless, some users say the label still appears after they use the tool. For other users, photos edited with the same tool have not been labeled at all.    The incorrect label also appeared on photos posted on Instagram by singer Halsey’s About Face cosmetics brand. The brand said the images were taken with an iPhone, lightly edited only in a photo app and did not use generative AI. In another test, an image specially modified to make it more difficult to train AI models was labeled. The unmodified version of the same image remained unlabeled.    During tests, images modified or fully generated using Canva, Photoshop, Adobe Firefly, Google Gemini and Apple Intelligence were not labeled. Although some files contained C2PA metadata showing their origin and an invisible marker called SynthID, Instagram did not recognize them. The label appeared only on images edited or generated in the Meta AI app. Images posted from a new account remained public for approximately 2 weeks, but the system did not classify them as AI content. It therefore remains unclear which indicators Meta uses to add the label.