Instagram's AI Content label is misfiring in both directions. Meta's system is stamping the tag on original photos that used nothing more than a background-removal tool, while fully AI-generated images posted with C2PA and SynthID watermarks are moving through the feed unlabeled. This is the second time in roughly two years the detection system has broken this way — Meta ran into an almost identical problem with its 'Made by AI' tag in February 2024.
The pattern users are reporting centers on Canva. Multiple Threads users say the AI Content label appears on any photo touched by Canva's Background Remover, even when the underlying image is a straightforward photograph. One user reported that removing a single speckle from an image was enough to get the entire photo flagged as AI on Instagram.
Canva has since told content strategist Jess Bruno that some of its assistive AI tools 'were being tagged as generative,' and says the tagging is now corrected. Canva's background removal help page states the tool 'doesn't add Canva's AI-generated content metadata to your design.' Threads users report mixed results after the fix — some Canva-edited images are still being tagged.
“every time there is bg remover involved”— Threads user, Instagram user report
Key facts
- 01Instagram's AI Content label is being auto-applied to photos edited with Canva's Background Remover, a machine-learning tool, not a generative one.
- 02Fully AI-generated images made in Google Gemini with C2PA and SynthID metadata embedded were not tagged by Instagram in tests.
- 03The only reliable trigger for Instagram's AI Content label in testing was content created in Meta's own AI app.
- 04Meta ran into a nearly identical labeling problem in February 2024 with its 'Made by AI' tag on Adobe-edited photos.
- 05Test images have been live on a public Instagram account for almost two weeks with no AI tag applied, despite Meta's crackdown on AI accounts announced last week.
The distinction Meta appears to be missing is the one between assistive machine learning and generative AI. Background removal and object selection have used machine learning inside tools like Photoshop for over a decade. Those features do not synthesize new pixels from a prompt, and they are not what the AI Content label was built to flag.
Meta has been vague about its detection method since launch. In 2024 the company said it would scan for IPTC and C2PA metadata to identify AI-generated or AI-manipulated content, and promised to tune the system to better reflect 'the amount of AI used in an image.' It has not published updated criteria, and did not respond to a request for clarification on the current system.
The false positives are hitting brands as well as individual users. About Face, the cosmetics company founded by singer Halsey, had a recent Instagram post auto-tagged with the AI Content label. The brand's social manager said no generative AI was used and pointed to standard editing inside Apple's Photos app on an iPhone.
“This photo was taken on my iPhone and then slightly edited by myself in the photo app. Our brand uses real artist[s] and real people to create everything you see across all our platforms”— About Face social manager, Cosmetics brand founded by Halsey
Only a narrow set of iPhone editing features should trigger a generative label — Spatial Reframing, Extend, and the updated Clean Up added in iOS 27. Those Apple Intelligence tools embed Google's SynthID watermark. According to Google's Gemini-based verification, the About Face image did not contain SynthID, which makes Meta's basis for the tag unclear.
The false negatives are the more damaging half of the failure. The Verge's Jess Weatherbed posted images fully generated in Google's Nano Banana model inside Gemini — depicting scenes and clothing she had never worn — and both images carried C2PA and SynthID signals. Instagram tagged neither. In the same test batch, images edited with Canva's Background Remover, Photoshop's background eraser, and Adobe Firefly's generative features also went untagged.
The only content that reliably triggered the AI Content label was material created inside Meta's own AI app. That result held even on a brand-new Instagram account posting AI images in quick succession — behavior that should look like an AI content farm to any moderation system. The account was not flagged, and this came after Meta announced a crackdown on AI-generated accounts last week. The test images have been live and public for almost two weeks.
There is a defensible reason for Meta to keep its detection heuristics private — publishing them helps bad actors evade the label. That defense weakens when the system tags a phone photo touched by a background eraser and misses a fully synthetic portrait carrying industry-standard watermarks. The signal-to-noise ratio of the label is now low enough that a viewer cannot use it to decide whether an image is real.
The commercial stakes for Meta are the credibility of the tag itself. Instagram sold the AI Content label as a trust primitive at a moment when generative image tools are cheap and abundant, and the label only works if it maps to something reliable. Right now it maps to Meta AI usage plus noise, which puts the company in the awkward position of having built a labeling system whose most consistent function is marking its own product. Until Meta names the metadata it actually scans for and treats C2PA and SynthID as first-class signals from other vendors, the label will keep training users to ignore it — the exact opposite of its purpose.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
