How to Remove AI Info from an Instagram Post (What It Actually Reads)

Instagram's AI info label reads C2PA and the IPTC digital source type. What triggers it, what a metadata clean actually fixes, and what it cannot.

Oct 26, 2026

AI Metadata Ninja Team

Instagram's "AI info" label is not a mystery toggle and it is not a ban. It is a small, boring piece of metadata that Meta reads when you upload, and once you know which field it reads, you can decide what to do about it.

One thing has to come first, because it decides whether any of this helps: the label is attached at upload time, and it cannot be edited off a post that is already published. Cleaning a file changes what happens on the next upload. If you want the label gone from a specific post, the only route is to delete that post and upload the cleaned file again.

That is the honest version. The rest of this is what "AI info" actually reads, which parts of it you control, and which parts you do not.

What the AI info label actually reads

Meta shipped a label called "Made with AI" in 2024 and renamed it "AI info" that July; the same label also renders as "AI content" on some posts. Meta's own documentation names the signals it uses, and they come down to three.

C2PA Content Credentials. A signed manifest — the c2pa chunk or APP11 marker — naming the generator, the actions taken on the file, and the ingredients. Firefly, Photoshop's Generative Fill, ChatGPT, Google Imagen and Microsoft Designer all write one.

The IPTC digital source type. A single XMP property, Iptc4xmpExt:DigitalSourceType, holding a vocabulary URI. The value that matters is trainedAlgorithmicMedia — wholly generated. Its sibling compositeWithTrainedAlgorithmicMedia means "a real photo with generated parts". This field is unsigned: any tool can add it, any tool can remove it. Midjourney writes it even though Midjourney does not sign C2PA.

Self-disclosure. You can tell Instagram the content is AI-generated yourself, and Meta's policy has always put that first.

Why the label appears on photos you did not generate

This is the case that catches people, and it is worth walking through because it is where the IPTC field earns its keep.

You photograph a sunset. You open it in Photoshop and use Generative Fill to add a streetlight. You export. The pixels are almost entirely yours, but the export now carries compositeWithTrainedAlgorithmicMedia — and Meta reads that field, so the post gets labelled.

The same thing happens to a clip shot on a phone and then run through an upscale or colour model that writes XMP, and to a scan that passed through a document-scanning app whose template embeds a digital source type.

And then there are the false positives, which are the reason posts like this one exist. A camera file that carries a signed C2PA manifest is not flagged, because the manifest's content is what matters, not its presence. A file with no metadata at all can still be labelled, because Meta also runs its own classifiers, and the invisible watermarking layers described further down.

The early-2026 test that should change your advice

Most "strip your C2PA" guides treat C2PA as the one signal that matters. An independent test circulated in early 2026 found the opposite order of importance: Instagram reacted to the IPTC digital source type, and did not react to a C2PA credential on its own.

That test is not Meta's documentation, and one sample is not a policy — treat it as a data point, not a verdict. But it is consistent with how the two fields differ. The IPTC field states outright that the content is AI-generated. A C2PA manifest states which tools touched a file, and that list includes ordinary editing.

The practical read: check the IPTC field first. If the file carries trainedAlgorithmicMedia or compositeWithTrainedAlgorithmicMedia anywhere in its XMP, that is the thing to remove.

The Reel cover thumbnail trap

If the label is on a Reel rather than a photo, look at the cover image before you look at the Reel.

Instagram uploads a Reel's cover thumbnail as its own file. That file is a JPEG or PNG, it is uploaded separately, and it can carry its own metadata — including an IPTC digital source type inherited from whatever you made the cover in. A Reel whose video file is perfectly clean can still show a label, because the cover is the file that tripped it.

The cover is also the easiest thing to fix. It is a still image, not a video, and a still image is what this tool handles best.

What a metadata clean can and cannot do

It can remove:

the C2PA credential, when the file carries one
the IPTC digital source type, and the rest of the XMP packet that holds it
EXIF fields naming the generator that wrote the file — Midjourney, Adobe Firefly, DALL·E, FLUX, ComfyUI, Canva, Leonardo, Ideogram, Stable Diffusion
the prompt and workflow records Stable Diffusion and ComfyUI leave in PNG text chunks

It cannot remove:

SynthID, Google's invisible watermark, which lives in the pixels
Meta's own Content Seal, which also lives in the pixels
Meta's classifier verdicts, which are not derived from your file at all
a label already attached to a published post

Those pixel-level watermarking layers are a different layer of the file entirely. They survive any metadata cleanup, which makes the honest answer to "will cleaning make the label go away" sometimes, for a metadata trigger, on the next upload — and never for a watermark trigger.

Cleaning does not discharge the disclosure duty

Removing metadata is a file operation. Disclosure is an obligation, and the two are not the same thing.

EU AI Act Article 50 transparency obligations have applied since 2 August 2026. Providers of systems generating synthetic images have to mark outputs in a machine-readable, detectable form, and deployers publishing deepfakes or AI-generated text have to disclose it. A cleaner that removes the machine-readable marker does not remove the obligation — it removes the evidence.

Meta's policy applies alongside all of that. And in August 2026 Instagram widened its own disclosure requirement: the "AI creator" label became "AI-generated profile" on 31 August 2026, and a profile featuring an AI-generated person without that label is now non-recommendable to non-followers.

There is a second Instagram label worth knowing about, because it is the one most people hit first: content that was only edited with AI is not labelled on the post itself but carries a label in the post's menu. That is the compositeWithTrainedAlgorithmicMedia case.

About the reach numbers

You will see figures online. Posts claim an AI label costs 15% of reach, or 30%, or 80%. None of those numbers come from Meta. Meta has published no reach penalty for a per-post AI label.

What Meta has published is the profile-level rule described above — a restriction on recommendation to non-followers, for profiles featuring an AI-generated person, and unrelated to metadata.

Treat the reach numbers as marketing copy. They are the standard sales device for "AI label removal" services, and they appear on exactly the pages whose business is selling the removal.

How to remove AI metadata from a photo before you upload it

Check the file first. Drop the JPEG into the free scanner. It names every AI trace it finds, which carrier holds it, and which generator wrote it. Nothing is removed yet.
Clean it. One click drops the metadata segments carrying the trace and writes a new file. The pixel data is byte-identical and the ICC colour profile survives.
Verify. The cleaned file is re-scanned, and you get the proof: what went, the byte count, and zero AI traces remaining.
Upload the cleaned file, not the original. This is the step people miss. Clean the file, then upload that file.

Do it at the last possible moment. Photoshop, Lightroom, Canva and CapCut all rewrite XMP or manifests on export, so a file cleaned on Monday can be dirty again by Wednesday's export. Cleaning is the last step of the pipeline, not the first.

If the file is a photograph rather than a render, the remove AI metadata from a photo guide covers what a clean costs you — GPS, camera body and lens live in the same EXIF block as the generator's name, so a block-level cleaner cannot remove one without the other.

Frequently asked questions

Can I remove the AI info label from a post I already published?

No. The label is applied at upload, and there is no edit control that removes it from a published post. Delete the post and upload the cleaned file again.

Does the AI info label affect my reach?

No figure from Meta supports the per-post reach claims circulating online. What Meta has published is a profile-level restriction: a profile featuring an AI-generated person without the "AI-generated profile" label is not recommended to non-followers.

Why did Instagram label a photo I took myself?

Usually an AI edit wrote compositeWithTrainedAlgorithmicMedia into the XMP. It can also be a false positive from Meta's own classifiers, which do not depend on your file's metadata at all.

Will removing AI metadata remove the invisible watermark?

No. SynthID and Meta's Content Seal are pixel-level. They are not metadata, and no metadata cleaner can touch them.

Does cleaning make me compliant with the EU AI Act?

No. Article 50 transparency obligations have applied since 2 August 2026. Removing the machine-readable marker does not discharge the disclosure duty — it removes the evidence of it.

Is there a setting in the app that turns the label off?

No. The label is metadata-driven. Your control is the content of the file you upload, plus the self-disclosure option where it applies.

Sources

Clean the file before you upload

Drop the image into the free AI metadata scanner. You get the report first — exactly which fields are there, which carrier holds them, and which generator wrote them — before anything is removed.