What the AI info label on Facebook and Instagram is built from
The AI info label is Meta's note on a post that its systems decided AI was involved. Meta applies it in two documented cases: when its systems detect an industry standard signal that content was created with AI, or when the person posting disclosed it. The wording has been "AI info" since July 1, 2024, when Meta renamed the earlier "Made with AI" label after complaints that slight AI retouches were being labeled, as the announcement at https://about.fb.com/news/2024/04/metas-approach-to-labeling-ai-generated-content-and-manipulated-media/ records. The label describes a signal or a disclosure. It does not certify that an image is real or fake. Three layers can carry the signal, and each one behaves differently when you edit the file.
The visible layer is the mark you can see. Meta's newsroom post from February 2024 says photorealistic images created with Meta AI have carried an "Imagined with AI" label since the tool launched. The Meta Help Center describes the same surface from the product side: content created or edited with Meta's AI tools and shared to Facebook, Instagram or Threads may be labeled AI content, and in some cases it carries a visible watermark. The wording on this page follows the label that most people see on Meta AI images. A visible mark is drawn into the picture, so it belongs to the pixels and travels with every copy, cropped or not. The February post is at https://about.fb.com/news/2024/02/labeling-ai-generated-images-on-facebook-instagram-and-threads/, and the help page behind these details is at https://www.meta.com/help/artificial-intelligence/1783222608822690/.
The metadata layer is the machine-readable declaration. The same help page says content made with third-party tools may carry an industry standard signal, and that Meta's systems read those signals to decide whether a label applies. The February 2024 newsroom post names the standards: Meta said it was building tools to read the "AI generated" information in the C2PA and IPTC technical standards. That is the layer this page works on. It is also the only layer a file editor can rewrite without touching the picture.
The invisible layer is the watermark inside the pixels. Meta's February 2024 post says the company puts visible markers on Meta AI images, along with both invisible watermarks and metadata embedded in the image files. It adds that using the two together improves how well the markers survive editing and helps other platforms identify them. An invisible watermark is a pattern in the image data, not a field in a file header. A tool that rewrites headers never reaches it.
The metadata layer this tool removes
C2PA Content Credentials are the larger record. The C2PA technical specification at https://spec.c2pa.org/specifications/specifications/2.5/index.html describes a signed manifest store that records how a file was made or changed, which tool signed it, and which assertions the signer made. On an AI export an assertion can state that a model produced the image. The manifest usually sits in a JUMBF box inside the image data, and an XMP packet in the file can point back to it. Taking the record out means taking the box, the pointer and the signed claim together.
IPTC fields are the smaller declaration. The Digital Source Type vocabulary at https://cv.iptc.org/newscodes/digitalsourcetype/ defines controlled values such as trainedAlgorithmicMedia for model output and compositeWithTrainedAlgorithmicMedia for a generated element added to a captured photo. The February 2024 Meta post calls the markers used for Meta AI images IPTC metadata and invisible watermarks, which puts the IPTC field on Meta's own published list of signals. The field is small, and it is one of the clearest statements a file can make about where it came from.
Generator and prompt fields are the plain text remainder. XMP often names the creator tool and the model, and some editors store the prompt and the negative prompt there as well. EXIF can carry a software tag with the exporter name. The strip treats the EXIF side narrowly: it removes Software, Make, Artist and ImageDescription tags only when the value names a known AI generator, and a tag from an ordinary camera or a non-AI editor can stay. None of these fields is signed, and any metadata viewer can read them. They tell a person which tool made the file even when no platform is involved, which is why the AI-named ones are worth clearing for their own sake.
The strip runs server side and rewrites the file once. It removes the C2PA manifest store and its JUMBF wrapper, the XMP reference that points at the manifest, the IPTC source type and the generator fields that name an AI tool. The pixels are copied through, so resolution and framing come back as they went in. The pass does not touch the visible mark or the pixel watermark, because both of those live in the picture rather than in the metadata block.
Why the visible mark and the invisible watermark stay
This page runs one operation, and that operation reads and writes metadata. A visible "Imagined with AI" mark is not metadata. It is drawn into the picture, so the clean copy still shows it exactly where it was. Taking a mark out of the pixels would mean rebuilding the image underneath it, which is a different operation and a different tool. Nothing in the metadata pass changes a pixel, so nothing in it can take that mark out of the picture.
Meta's invisible watermark is in the same position, for a different reason. The February 2024 post says the watermark is embedded in the image file alongside the metadata, and that using both improves how well the markers last. A watermark is meant to survive the edits that break metadata: re-saving, compression, cropping and format changes. A metadata strip leaves the pixel data alone, so the watermark stays with it. Meta does not publish the decoder for that watermark, so a metadata viewer cannot tell you whether a given image carries one.
The result is a file that declares nothing and still looks the same. The declarations are gone, and the picture is untouched. Meta's systems may still read the signals that remain, including the visible mark, the invisible watermark and whatever else the platform checks. The help page states the limit from its side as well: not all AI content contains the information needed to identify it. The reverse holds here, because a signal outside the metadata block can survive a clean.
That leaves a practical rule. If the visible mark is the problem, this is not the tool for that layer, and no metadata pass will be. If the metadata is the part you want gone, this page handles it and says so plainly. Read the two layers separately, and neither promise gets blurred.
How it works
How to strip the AI metadata before you post
Step 1: Export the final image
Set the crop, color and lettering before the final export. An editor that has Content Credentials turned on will sign the file every time it saves, so cleaning has to come last, after all edits and immediately before upload. When possible, begin with the original export; a screenshot has already lost the metadata and picked up compression damage.
Tip
If your editor can export with Content Credentials off, turn it off at the source and save the strip for files that already carry a record.
Step 2: Run the metadata strip
Drop a PNG, JPEG or WebP into the panel and start the job. The pass removes the C2PA manifest store, its JUMBF wrapper, the XMP pointer to the manifest, the IPTC source type and the generator fields that name an AI tool. The image is copied through at its original size, so the result is the same picture with a smaller metadata block. iPhone photos often arrive as HEIC, so convert those to JPEG or PNG before you upload.
Tip
The visible mark and the invisible watermark are not part of this pass. If the file carries either one, the clean copy carries it too.
Step 3: Download the clean copy and post it
Download the cleaned file and upload that copy, not the original in your camera roll. If you resize or convert the image afterwards, use an editor that does not re-embed AI metadata on export. Meta does not require a label for images under its current rule, though images may still receive one when its systems detect AI. Check the disclosure options in the composer and make that call yourself.
Tip
The strip changes the file. It does not change what you write in the post, and a disclosure you owe is still yours to make.
What comes off, and what stays
C2PA manifest store
The signed record that can state a model made the image. The strip removes the manifest store, its JUMBF wrapper and the XMP pointer that leads back to it, so the clean copy carries no signed claim.
IPTC Digital Source Type
The field that names media as trained algorithmic output or a composite. The strip removes it with the rest of the AI-related IPTC block.
XMP and EXIF generator fields
The tool name and the model name and, in some editors, the prompt. The pass removes the XMP and EXIF fields whose values name an AI generator. A standard camera or editor tag can stay in the file.
The visible mark stays
The "Imagined with AI" corner mark is drawn into the pixels. The strip copies the pixels through, so the mark stays in the picture and the clean copy still shows it.
Meta's invisible watermark stays
Meta embeds an invisible watermark in Meta AI images alongside the metadata. The pattern lives in the pixel data and is built to survive edits, so a metadata strip does not remove it.
The label decision stays with Meta
Meta applies the AI info label on its own side, after upload, from the signals its systems can read. No file edit on this page makes that decision.
Where this page stops
FAQ
Frequently asked questions
Short answers about the Meta AI metadata remover, from the signals Meta documents to the two layers this tool leaves in the picture.
Explore more tools
Remove AI Label
The cross-platform page on the markers platforms read at upload and the strip that clears them.
Remove AI Info from Instagram Posts
The step-by-step guide to the Instagram side: why the label appears and how to clean the file before you post.
Facebook AI Label Remover
The Facebook guide to the label, the options menu and the metadata strip that keeps the next upload clean.
Image Metadata Viewer
See which C2PA and IPTC fields a Meta AI image carries before the strip.
C2PA Remover
The credential layer on its own, for times when only the signed manifest is the problem.
All tools
Every remover and detector on one page, grouped by what each one cleans or checks.
Strip the AI metadata before the upload
One pass removes the C2PA manifest, the IPTC source type and the AI generator fields while the pixels come back unchanged. The visible mark and the invisible watermark stay, exactly as the boundary at the top of this page says.