Instagram can show an "AI info" label on a post when the file you upload carries machine-readable markers that say the image was made or edited with AI. That label is separate from EXIF and GPS. Instagram re-encodes what viewers download and strips camera metadata from that public copy, but Meta still reads provenance fields from your original upload to decide whether to apply the label. If you want the post without that badge, the file you send has to arrive without the markers Meta looks for, which usually means removing C2PA Content Credentials and the IPTC Digital Source Type field before upload.
What you see on Instagram
Meta's wording has shifted ("Made with AI" became "AI info" in July 2024), but the pattern is the same: a notice that the content was created or modified with AI. Since September 2024, content Meta judges to be only edited with AI gets the label in the post's menu; content it detects as generated keeps the label on the post. It can appear on posts, stories and reels. The label is not written into EXIF the way GPS is. It is a platform decision at ingest time.
That distinction trips people up. You can download someone else's Instagram photo and find no EXIF at all, then assume metadata never mattered. For the AI label, it did matter on the way in, even if the served JPEG is clean.
What Meta reads from your file
Meta says it reads "AI generated" information in C2PA and IPTC metadata to apply its AI info label on Facebook, Instagram and Threads. In practice that means two things in the file bytes:
- A C2PA manifest (Content Credentials), stored as JUMBF in JPEG APP11, a
caBXchunk in PNG, or aC2PAchunk in WebP. Generators such as ChatGPT, Adobe Firefly, and others embed a signed record whose first action is oftenc2pa.createdwith a digital source type of trained algorithmic media. The full spec is in what C2PA Content Credentials are.
- IPTC Digital Source Type in XMP, tag
XMP-iptcExt:DigitalSourceType, often the URIhttp://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMediafor wholly generated images. That field is unsigned: anyone can add or delete it without a signature. See Digital Source Type for the vocabulary.
Some files carry both. Meta's public statements refer to both families. Meta can change detection over time; treat vendor documentation as the source of truth, not this page.
OpenAI and Google also describe invisible pixel watermarks (SynthID and similar) that are not metadata at all. For files made with other companies' tools, Meta relies on C2PA and IPTC markers in the file, alongside self-disclosure by the person posting and its own detection. Watermarks are covered under limits below.
Instagram strips EXIF but still labels AI content
These two behaviors coexist.
| What happens | EXIF / GPS on the public copy | AI info label on the post |
|---|---|---|
| You upload a phone JPEG with GPS | Stripped from re-encoded copies viewers save | Usually no, unless other markers are present |
| You upload a ChatGPT PNG with C2PA | Also stripped from Instagram's served JPEG | Often yes, if Meta read C2PA or DigitalSourceType at upload |
The does social media remove metadata guide walks through platform-by-platform stripping. Instagram removes EXIF from what it serves. It does not follow that removing EXIF before upload stops an AI label when C2PA or DigitalSourceType is still in the file you upload.
Meta's privacy policy still lists metadata about uploaded content among what it collects. Cleaning the file on your device before upload is the only way to keep those provenance fields off Meta's servers.
Check whether your image carries AI markers
On a computer with ExifTool installed:
exiftool -G1 -a -s -JUMBF:all photo.jpg
exiftool -G1 -a -s -XMP-iptcExt:DigitalSourceType photo.jpgA JUMBF block labeled c2pa means a manifest store is present. A DigitalSourceType line with trainedAlgorithmicMedia or compositeWithTrainedAlgorithmicMedia flags generative content in plain XMP. A manifest can also be stored remotely and referenced from XMP as dcterms:provenance; exiftool -G1 -a -s -XMP:all photo.jpg lists it as [XMP-dcterms] Provenance.
For validation detail, c2patool photo.jpg --info reports manifest count and validation status. Uploading to contentcredentials.org/verify is another option.
The EXIF reader on this site has an AI and provenance panel: it shows whether a C2PA manifest is present, which app wrote it, and the DigitalSourceType declared in the manifest or in XMP. It does not validate the signature; c2patool or Content Credentials Verify do.
Remove C2PA and DigitalSourceType before you post
With the metadata remover
The metadata remover accepts JPEG, PNG, TIFF and WebP and returns the cleaned file. Two of its options matter here:
- All metadata removes the C2PA manifest, DigitalSourceType, EXIF, IPTC, XMP and the embedded thumbnail. It keeps the Orientation tag and the ICC profile unless you untick that option.
- AI markers only removes the C2PA manifest, DigitalSourceType and the
dcterms:provenancelink to a remote manifest. Camera EXIF, dates and GPS stay.
Workflow:
- Drop the image into the EXIF reader. Its AI and provenance panel shows whether the file carries a C2PA manifest or a Digital Source Type.
- Open the metadata remover, upload the image and pick one of the two options.
- Upload the downloaded file to Instagram, not the original.
Processing is single-file, no account, and the upload is deleted after the response.
Keep camera EXIF, drop only C2PA (ExifTool locally)
If you want to keep date and camera tags but remove the manifest:
exiftool -JUMBF:all= photo.jpgTo clear only DigitalSourceType while leaving other XMP:
exiftool -XMP-iptcExt:DigitalSourceType= photo.jpgThe AI markers only option of the web remover runs both, plus -XMP-dcterms:all= for the remote manifest link. More background in how to remove metadata from photos.
What removing metadata does not do
Stripping C2PA and DigitalSourceType removes the fields Meta cites for the AI info label. It does not change the pixels.
- Invisible watermarks such as SynthID stay in the image data. No metadata tool removes them. OpenAI documents SynthID on generated images alongside C2PA.
- Re-exporting through an AI tool or a C2PA-aware editor can embed a fresh manifest. You would need to strip again before the next upload.
- Absence of credentials is not proof the image is a camera original. Most photos on the web carry no C2PA at all. Removing metadata only removes the label's machine-readable trigger, not the history of how the picture was made.
- Generator terms can forbid it. Adobe's generative AI terms prohibit removing Content Credentials from Firefly output, and OpenAI's terms of use forbid presenting output as human-generated when it was not.
- Platform policy can add labels by other means later. This page describes file metadata as Meta describes it today, not a guarantee about every future post.
There are legitimate reasons to strip provenance too: manifests can include your name, edit history, and thumbnails of source images you may not want to publish.