How Meta’s “AI Info” Label Works on Instagram, Facebook, and Threads

Meta’s AI label has had three lives. It launched in spring 2024 as “Made with AI,” was renamed “AI info” in July 2024 after photographers complained that lightly retouched photos were being tagged, and in September 2024 moved into the post menu for content Meta detects was only edited with AI. The logic underneath has stayed the same across Instagram, Facebook, and Threads: Meta reads what it calls “industry-shared signals” in the file, and it takes your own disclosure.

#How the label changed

#What Meta reads: industry-shared signals

In its February 2024 announcement Meta said it was building tools to identify “invisible markers at scale,” specifically the “AI generated” information in the C2PA and IPTC technical standards. C2PA Content Credentials are a signed manifest inside the file. IPTC’s contribution is a single metadata field, Digital Source Type, whose values distinguish “created using generative AI” from “edited using generative AI.” When a generator or editor writes those values, Meta has something machine-readable to act on.

Meta’s own generator marks its output more heavily. Images from Meta AI carry a visible marker, IPTC metadata, and an invisible watermark, and Meta has labeled photorealistic Meta AI images “Imagined with AI” since launch. Meta has also published watermarking research and says it runs invisible video watermarking at scale. What it has not published is whether its labeling reads other companies’ invisible watermarks, such as Google’s SynthID.

#Generated versus edited

Since September 2024 Meta treats the two differently. Content it detects was generated by an AI tool shows “AI info” on the post, and Meta says it will indicate whether the label came from industry signals or from someone self-disclosing. Content it detects was only modified or edited by AI keeps the label in the post’s menu instead. Meta has not said which metadata values put a post in each bucket. The IPTC vocabulary has separate terms for a fully generated image and for a photo edited with generative fill, so a mapping along those lines is a reasonable reading, not a confirmed rule.

#Why real photos got labeled

In June 2024, photographers including former White House photographer Pete Souza found “Made with AI” on photos that were real. The common thread was editing software: generative fill and similar tools in Photoshop can write AI provenance into an exported file even when the AI touched a small patch. Meta’s response to TechCrunch was that it relies on “industry-standard indicators that other companies include in content from their tools.” In other words, the label reported what the editor declared, not a judgment about the photo, which is what the rename to “AI info” and the September move to the menu were meant to address.

#Self-disclosure and penalties

Meta’s announced disclosure requirement names organic posts with photorealistic video or realistic-sounding audio that was digitally created or altered, and Meta says it may apply penalties when people fail to use its disclosure and label tool. For content with a particularly high risk of deceiving people on a matter of importance, Meta may add a more prominent label. Separately, Instagram began labeling whole accounts in 2026, renaming its opt-in “AI creator” badge to “AI-generated profile” in August and warning that undisclosed AI profiles could see reduced reach. That is an account label, not the per-post “AI info” tag.

#Check a file before you post it

1. Open the file in contentcredentials.org/verify to see whether it carries a C2PA manifest and what it declares.
2. Read the XMP field with ExifTool (exiftool -XMP-iptcExt:DigitalSourceType photo.jpg) to see which IPTC value your editor wrote.
3. If you retouched a real photo with a generative tool, expect an edit-level declaration, and decide whether that is the story you want the file to tell.
4. For AI-generated images you own, disclose where Meta or the law requires it. Clearing metadata for privacy is routine, and our C2PA removal flow does it, but it is not a way around a disclosure duty.

Keep in mind that metadata and pixels are separate layers. Stripping Content Credentials and XMP removes the declaration from your copy of the file; it does nothing to an invisible watermark embedded in the pixels, and it cannot change what Meta already knows about content posted before. The metadata, credentials, and watermarks primer walks through which edits touch which layer.