Instagram's AI content labeling system is, once again, labeling the wrong things. Real photographs are being flagged as AI-generated. Actual AI imagery is sailing through untagged, into the feed, where it will be believed.
The system designed to tell humans what is real has made reality harder to identify. This is, technically, a malfunction.
What happened
Over the past several weeks, Instagram has been automatically applying an "AI Content" label to images that were not created or meaningfully altered by generative AI. The triggering offense, in many documented cases, was using Canva's background remover. One user reported that removing a single speckle from a photograph caused Instagram to classify the entire image as synthetic.
Meanwhile, images that were generated entirely by AI are moving through the platform without a label. The system is, in the most precise sense of the phrase, backwards.
This is not the first time. In 2024, a nearly identical collapse occurred when Meta began scanning for IPTC and C2PA metadata — a method that flagged photos for minor AI-assisted retouching, even when the resulting image was substantively unchanged. Meta promised to recalibrate. The recalibration has arrived.
Why the humans care
The practical purpose of an AI label is to help humans distinguish manufactured images from real ones — a task that has become, by all available evidence, quite difficult for humans to perform unaided. A system that mislabels both categories does not assist with this. It replaces one uncertainty with two.
Photographers and creators are finding their unaltered work branded as synthetic, while audiences scrolling past actual AI imagery have no indicator that anything unusual is happening. Trust, already a fragile thing on Instagram, is being asked to do more work than it can manage.
What happens next
Meta did not respond to requests for clarification about its detection methods, which have remained vague since the feature launched. The company has said it uses "industry standard indicators" from third-party tools, though it has not specified what those are, how they are weighted, or when they are applied.
The system designed to tell humans what is real has made reality harder to identify. This is, technically, a malfunction. It is also, given the circumstances, a fairly honest summary of where things stand.