Instagram’s attempt to make synthetic imagery obvious to users is producing the opposite effect: visible labels that look unreliable and random.
The Verge reported that users have seen the system gone haywire, with an “AI Content” label applied to pictures that were not created with generative tools and missed on other images that were AI.
The mislabelling appears to have multiple triggers, according to the reporting: edits made with tools such as Canva’s Background Remover and even tiny blemish fixes have been enough to flip the tag on a photograph.
False positives
Those false positives are not new. Meta previously began scanning images for provenance metadata such as IPTC and C2PA signals after earlier controversy, and past versions of the label system swept up retouched photos whose Adobe metadata hinted at AI-assisted edits IPTC and C2PA metadata.
The Verge adds that Meta has remained vague about the precise mix of metadata checks and classifiers it uses, which makes the visible label feel like an unexplained black box vague about its methods.
At the same time, Instagram has been building a separate, account-level policy that requires profiles using AI to generate people to self-identify; accounts that do not self-label can be in reach on Reels and Explore, with an appeal route available required to self-label.
Shift
That shift places some of the disclosure burden on creators rather than on post-level detection, but it also raises the stakes for any detection system that can both over-label ordinary photos and miss synthetic ones.
The technical landscape helps explain why this keeps happening: platforms now combine provenance metadata, visual watermarks, and trained classifiers into a detection stack, and each method has limits.
Heavy-handed reliance on provenance fields treats editing tools’ metadata as a proxy for synthetic content, which can catch genuinely edited photographs as “AI” even when the change is trivial. Equally, visual classifiers still produce false negatives on convincing synthetic images.
Simple and practical
For people who use Instagram to share photographs - creators, journalists, families - the practical effect is simple. A tag that flags everyday edits as AI and lets some fakes slip through makes the label less helpful and harms trust in the platform’s signals.
The next concrete test will be how Meta handles appeals and whether the account-level self-labelling enforcement reduces the visible noise; until that is resolved, the label will read as noise, not information.
