Where the label comes from
Generative fill, object removal, canvas expansion and background replacement can make an editor record a standard signal describing the action. That signal does not necessarily claim that the entire photograph was generated. A C2PA history can identify an original ingredient, the editing application and a sequence of actions.
Instagram cannot infer the artistic context of every edit. It can use industry signals, a creator’s disclosure and other available indicators to decide where AI info appears. A portrait with one AI-assisted correction and a fully synthetic scene may therefore enter the same broad category even though their histories are different.
Meta changed the system after false impressions
In July 2024 Meta said that minor AI retouching could trigger a Made with AI label that did not align with people’s expectations. It renamed the label AI info, then moved information about content detected as only AI-edited into the post menu. More prominent treatment remained for content detected as generated.
That distinction matters. AI information is not automatically an accusation of deception. Meta says it generally keeps this content available with context unless it violates another policy. Separate demotion can apply when fact-checkers rate content false or altered; that is a different process.
What to inspect before publishing
Keep the camera original and the editor project. Inspect the exported copy for an active C2PA manifest, declared actions, claim generator and validation result. A Software EXIF tag is useful context but does not prove full generation by itself.
- Do not interpret “manifest absent” as “AI definitely not used.”
- Do not edit signed claim text manually; doing so breaks validation.
- If a platform gets it wrong, preserve the original, the inspection result and the export history.
Will stripping metadata fix it?
Re-encoding or cleaning can remove an embedded manifest, but it does not turn the output into proven human-made content. A platform may have other signals, while the absence of provenance can make it harder to show the chain from camera to final image. Diagnose the signal before deciding what a publication copy should retain.
What to do: the short answer
Do not delete the only copy and do not edit a signed claim by hand. Keep the camera original, editable project and final export, then inspect C2PA separately from EXIF/XMP. If AI only changed a background or performed a minor retouch, the edit history is the evidence that distinguishes AI-edited work from a fully generated scene.
- For privacy, create a separate publishing copy and remove only supported fields you do not need to share.
- For a false-label appeal, retain the source, inspection report and a screenshot of the platform decision.
- Do not assume a clean file guarantees removal of AI info; Instagram can use signals outside the file.