Instagram's AI labels flag real photos while verified AI images go untagged

Labels misfiring in both directions

Instagram's visible "AI Content" labels, meant to let viewers recognise synthetic media at a glance, are failing in opposite ways at once, according to The Verge. Users report the tag being applied automatically to photographs that were never created or edited with generative AI, while machine-generated imagery is being posted without any label whatsoever.

A repeat of 2024

This is not the first time. In 2024, a few months after labeling launched, a "Made by AI" tag caught photographs that had been retouched with Adobe's generative tools, even when the edits were so minor the picture was substantively unchanged. Meta said at the time it would adjust labeling to reflect "the amount of AI used in an image." It had announced in February 2024 that it would scan images for IPTC and C2PA metadata, which can indicate with reasonable confidence that generative AI was involved, but the company has stayed vague about its actual methods and did not respond to The Verge's request for clarification.

The Canva connection

Many of the recent false positives involve Canva's Background Remover. One Threads user told The Verge the label appeared "every time there is bg remover involved," and another said removing "a speckle" caused an entire photo to be flagged. Content strategist Jess Bruno raised the issue with Canva and was told some of the platform's assistive AI tools "were being tagged as generative" and now tag correctly. Canva did not respond to a request for comment.

The fix may not be complete. The Verge reports that Threads users still see the label on Background Remover edits after Canva's claimed resolution, while others found that images edited before the fix were never tagged and that the tool had not applied the C2PA metadata Meta supposedly detects.

Background removal does rely on machine learning, but it is the assistive variety that Photoshop has offered in selection and object-removal tools for over a decade — not the generative text-to-image systems associated with deceptive imagery.

Cases with no visible explanation

Some flags have no clear cause at all. The Verge describes a case in which Meta tagged an image that had been "poisoned" — subtly tweaked so it degrades AI models trained on it — while leaving the untouched version of the same content alone.

Recent photos from About Face, the cosmetics company founded by the singer Halsey, were also auto-tagged. The brand's social manager replied publicly that the pictures were taken on an iPhone and lightly edited in the Photos app, with no AI involved and real artists producing the brand's content. The handful of Apple tools that do add generative metadata — Spatial Reframing, Extend, and the updated Clean Up in iOS 27, which use Apple Intelligence and embed Google's SynthID watermark — apparently were not the trigger either: The Verge checked the tagged images with Google's Gemini-based verification and found no SynthID present.

Testing finds little consistent logic

In hands-on tests described by The Verge's Jess Weatherbed, images edited with Canva's Background Remover, Photoshop's background-erasing tool, Adobe Firefly, Google's Nano Banana model, and iOS Apple Intelligence features all went unlabeled on Instagram — even though several carried confirmed C2PA and SynthID signals. The only reliable trigger for the tag was content edited or fully generated in Meta's own AI app.

A brand-new account created with minimal information and posting images in quick succession — deliberately behaving like an AI farm — likewise received no image labels and was not flagged as an AI-generated profile, even after Meta publicly announced a crackdown on AI accounts the week before. The posts remained live and public for nearly two weeks.

Why it matters

Provenance labeling only works if people believe it. False positives punish ordinary photographers and brands who did nothing wrong, and risk pushing honest creators away from everyday editing tools. False negatives let genuinely deceptive synthetic media circulate with an implicit stamp of authenticity, since the absence of a label reads as a clean bill of health. Meta may keep its detection methods vague to stop bad actors from gaming the system, but with no discernible pattern behind the tags, users are left with a signal that manufactures doubt about real work while missing the synthetic content it was built to catch. The Verge's own testing ended with the reporter concluding she could no longer trust the labels themselves — which is precisely the outcome a labeling system exists to prevent.