Detect what triggers the AI info label
Meta does not look at your picture and decide it is fake. For most labeled posts it reads what is written inside the file. Pick a photo or video above and the report lists every signal it finds, each tagged AI (reads as an AI disclosure) or Info (provenance data that does not claim AI). That makes it a quick way to check an image for AI metadata before you post.
The detector looks for:
- C2PA Content Credentials, stored as a JUMBF block in JPEG, a caBX chunk in PNG, a C2PA chunk in WebP, or a uuid box in MP4, MOV, HEIC and AVIF.
- IPTC DigitalSourceType in XMP, with AI values such as trainedAlgorithmicMedia and compositeWithTrainedAlgorithmicMedia.
- Tool names like Firefly, Generative Fill, DALL-E, Midjourney, Gemini or Stable Diffusion in XMP CreatorTool, edit history, EXIF Software, ImageDescription, UserComment and Artist.
- Prompt chunks that Stable Diffusion and ComfyUI write into PNG files.
Clean the file and download
Press Clean & download and you get your-file-clean.jpg (or the same name with your own extension). The tool deletes the Content Credentials, the XMP and IPTC fields that carry the signal, and blanks AI tool names in EXIF text. It then scans the new file again so you can see the result for yourself.
Pixels, audio and video frames are never re-encoded, so quality does not change. EXIF orientation and ICC color profiles stay. In MP4, MOV, HEIC and AVIF files the removed boxes are swapped for empty ones of the same size, which means no chunk offsets move and the video keeps playing exactly as before.
If you also use the meta glasses photo converter or the video converter, clean the file first, convert second, then post. That is the easiest way to clean metadata before posting without losing the Ray-Ban Meta tags those tools add.
Why real photos get labeled AI info
The label was never meant to punish a normal photo, but the signals it relies on are blunt. A single AI-assisted brush stroke leaves the same kind of marker as a fully generated image, and the file does not record how big the edit was.
Photoshop Generative Fill, Remove and Expand
Adobe attaches Content Credentials to files that use its generative features, and the credentials can include an IPTC source type that says AI was used. Delete one power line from a sunset with Generative Fill and the export can carry the same flag as an image made from scratch. So yes, Generative Fill can trigger the label, even when 99 percent of the frame is a real capture.
Lightroom, Canva and background removers
Lightroom's Generative Remove works the same way. Canva's AI editing tools, Google's Magic Editor, phone gallery apps with generative erase, and some background removers that use generative models can also write markers. If your caption says nothing about AI and the label still shows up, open the report: the cause is usually a single field written by whatever app touched the file last.
What gets removed vs what can't be removed
Removed
- C2PA / Content Credentials manifests
- XMP packets and extended XMP that carry AI or provenance data
- IPTC DigitalSourceType and AI system fields
- Photoshop / IPTC blocks that name an AI tool
- AI tool names in EXIF text and JPEG comments
- Stable Diffusion and ComfyUI PNG text chunks
Cannot be removed
- Invisible pixel watermarks such as Google SynthID and Meta Video Seal
- Meta's own AI classifiers that analyse the image or video
- The "AI info" self-disclosure toggle you switch on while posting
- The label on a post that is already live
So treat this as a metadata cleaner, not a guarantee. It fixes the most common cause of wrongly labeled posts and nothing more.
Instagram, Facebook and Threads: how each reads the file
All three use one Meta system. According to Meta's labeling announcement, it checks for industry-standard indicators (C2PA and IPTC) and for people disclosing AI use themselves. Meta's help page explains what the label means. A clean file works for Facebook AI info removal and for Threads AI info the same way it does on Instagram.
The label itself has changed. "Made with AI" became "AI info" in July 2024, and since September 2024 AI-edited content shows it in the post menu while fully generated content keeps a visible label. For a Reel, a Story or a feed post, clean the finished MP4 or MOV, because a later re-export can write new Content Credentials. Phone photos in HEIC and AVIF are handled too.
If the label is already on a live post, deleting the metadata afterwards does nothing. Upload the clean copy as a new post. If you added the label yourself, switch the toggle off in the post settings. For a wrong label that you did not add, use Instagram's Help and report-a-problem options; Meta does not promise a result.
Privacy: nothing is uploaded
The scanner and the cleaner are a small piece of JavaScript that runs in your browser tab. Your photo or video is read into memory, rewritten there, and handed back as a download. No request carries your file anywhere, which also makes this a safe way to clean private metadata such as editing history and author fields. Our privacy policy covers the rest of the site.
When you should keep the label
Use this to fix wrongly labeled real photos and to clean private metadata. Content that was actually AI-generated should still be disclosed, as Meta's policy requires.
If you generated a realistic person, place or event with an AI model, leave the credentials in or use the disclosure toggle when you post. Hiding it puts your account at risk and misleads the people who follow it. Stripping the tag from a picture you actually took, or from an edit that was minor, is a different matter.
More detail: why real photos get flagged, fixing a post that is already live, and Photoshop Generative Fill and the AI label. To only inspect a file, use the AI metadata checker; to strip C2PA from any image or video, see remove Content Credentials.
Want the vocabulary behind the fields? The IPTC publishes the full list of digital source type values. You can also browse more free browser tools in the tools hub.
FAQ: removing AI info from Instagram posts
Why does Instagram say "AI info" on my real photo?
Meta reads industry-standard signals in the file, mainly C2PA Content Credentials and IPTC metadata. Editing tools write those signals even for small AI-assisted edits, such as removing a stray object with Generative Fill, so a mostly real photo can still carry an AI marker.
Does removing metadata remove the AI label?
If the label came from the file metadata, a clean copy is usually not labeled automatically when you post it. It cannot beat invisible watermarks or Meta's own classifiers, and there is no guarantee either way.
Can I remove AI info from a post that is already live?
Cleaning a file does not change a post that has already been published. Delete the post and upload the clean copy as a new one, or turn off the toggle in the post settings if you added the label yourself. Meta's help center lists the current ways to dispute a label, and results vary.
Is "AI info" the same as "Made with AI"?
Yes. Meta renamed "Made with AI" to "AI info" in July 2024 because real photos edited with AI tools were being caught. Since September 2024, content that was only edited with AI shows the label in the post menu, while fully AI-generated content keeps the more visible label.
Does it work on Facebook and Threads?
Yes. Instagram, Facebook and Threads use the same Meta system and read the same signals, so a file that is clean for one is clean for the others as far as metadata goes.
What is C2PA?
C2PA is an open standard for Content Credentials: a signed record of where a file came from and what was done to it. It is stored inside the file (a JUMBF block in JPEG, a caBX chunk in PNG, a uuid box in MP4 and HEIC). Cameras, editors and AI tools can all write it.
What is IPTC DigitalSourceType?
It is an IPTC metadata field that says how media was made. Values such as trainedAlgorithmicMedia and compositeWithTrainedAlgorithmicMedia mean generative AI was involved, while digitalCapture means a normal camera shot. It appears as a URL under cv.iptc.org/newscodes/digitalsourcetype/ or as the bare word.
Does it remove SynthID or other invisible watermarks?
No. SynthID, Meta Video Seal and similar watermarks live in the pixels or audio, not in metadata. No metadata tool can remove them without re-rendering the picture, and this one never touches your pixels.
Is my file uploaded?
No. The file is read and rewritten inside your browser tab with plain JavaScript. Nothing is sent to a server, and the page keeps working if you go offline after it has loaded.
Is it okay to remove the label from AI-generated images?
Use the tool for real photos that were mislabeled and to strip private metadata. Content that is actually AI-generated and looks realistic should still be disclosed, as Meta's policy requires.
Can I remove AI info from a Reel or a Story?
Yes, clean the MP4 or MOV before you upload it. Reels, Stories and feed posts are read the same way. Re-exporting the clip in an editor that adds Content Credentials will put the signals back, so clean the final export.
Does it work on HEIC photos and MP4 videos?
Yes. HEIC, AVIF, MP4 and MOV files are cleaned by replacing the offending boxes with empty ones of the same size, so the video and image data stay exactly where they were and nothing is re-encoded.
Does Photoshop Generative Fill or Lightroom Generative Remove trigger the label?
They can. Adobe's generative features can write Content Credentials and an IPTC source type to the exported file, and Meta reads both. Whether a post is labeled also depends on the export settings and on what Meta's other checks find.