Which Platforms Label AI Images, and How
Five platforms, four different mechanisms. The one a platform uses decides whether cleaning a file before upload changes anything at all.
The mechanism is the whole story
Platforms are commonly described as detecting AI images, as though they all do the same thing. They do not, and the differences decide what any action on your part can achieve.
There are four distinct approaches in use, and a platform may combine them:
- Reading a C2PA manifest — a signed provenance record written into the file by the editing or generation tool.
- Reading IPTC fields — older, simpler text metadata, including a field that states the digital source type.
- Running a classifier over the pixels — a statistical guess, entirely independent of any metadata.
- Requiring the uploader to disclose — a policy obligation on the creator rather than any technical check.
Meta: Instagram and Facebook
Meta reads the C2PA manifest. If a file arrives carrying Content Credentials that name a generative tool, the post gets an AI info entry.
This is metadata-driven, so a file with no manifest gives Meta nothing to read. It is also why real photographs get labelled after a single retouching stroke made with a generative feature: the manifest records that such a feature was used, not how much of the image it produced.
LinkedIn also reads C2PA, and displays a small CR badge on images carrying credentials. The badge is clickable, and it reveals more than Meta shows — the tool involved and the party that signed the credential.
For anyone posting client work, that is worth knowing before you upload rather than after. The badge does not merely say that AI was involved; it names the software.
TikTok
TikTok combines both approaches. It reads C2PA credentials to label synthetic media automatically, and it separately requires creators to disclose realistic AI-generated content depicting people or scenes.
The automatic reading applies whether or not the creator discloses. The disclosure requirement is a policy obligation that applies whether or not any credential is present, and failing it is a rules problem rather than a metadata problem.
Pinterest: the one that behaves differently
Pinterest is the exception worth understanding, because it does not fit the pattern the others set.
It reads IPTC metadata rather than C2PA, and has shown an AI modified label on pins since April 2025. IPTC is a much older and simpler standard than C2PA, and the relevant field is a plain text statement about the digital source rather than a signed manifest.
More importantly, Pinterest supplements that with its own classifiers for content that arrives unmarked. That second layer reads pixels, not metadata — which means Pinterest is the platform where cleaning a file is least likely to be sufficient. Removing the IPTC field removes one input, and the classifier is still there.
YouTube
YouTube took the policy route. Its disclosure rules, introduced in 2024 and enforced from early 2025, require creators to label realistic altered or synthetic content that could mislead viewers about real events, people or places.
That is an obligation on you rather than a check on the file. No amount of metadata work interacts with it, because nothing is reading the metadata for this purpose. The relevant question is whether the content is the kind the policy covers.
What cleaning a file actually changes, platform by platform
Putting the mechanisms together gives a straightforward answer to a question that otherwise seems to have five different answers.
- Meta and LinkedIn read a manifest, so a file with no manifest produces no label. Removal is effective, but only before upload — a published post keeps the label it was given.
- TikTok reads credentials the same way, but the disclosure requirement is independent of the file and unaffected by anything you remove.
- Pinterest reads IPTC and also runs classifiers. Removal addresses one and not the other.
- YouTube reads nothing. Its rule is a policy about the content itself.
The asymmetry nobody designed on purpose
Because these systems key on metadata that the generating tool chooses to write, they catch the tools that participate and miss the ones that do not.
Adobe writes credentials, so a photographer who removed a distraction in Photoshop gets labelled. As of early 2026 Midjourney does not embed C2PA credentials, so a fully generated image from that tool arrives with nothing for a platform to read.
The result is a labelling system that lands hardest on the workflows closest to ordinary photography, and lightest on some of the fully generative ones. That is not an argument against provenance metadata, which does something genuinely useful. It is a reason to treat the presence or absence of a label as much weaker evidence than it appears to be.
See what a file tells a platform
See exactly which AI markers an image carries before you publish it. This tool only reads — nothing is changed, and nothing is uploaded.
Open AI Metadata CheckerFrequently asked questions
Which platforms actually read metadata to detect AI images?
Meta, LinkedIn and TikTok read C2PA Content Credentials. Pinterest reads IPTC metadata and additionally runs its own pixel classifiers. YouTube does not read metadata for this purpose at all — its approach is a disclosure rule placed on the creator.
Does removing metadata stop every platform labelling my image?
No. It removes the trigger on platforms that read a manifest, such as Meta and LinkedIn. It does not affect Pinterest classifiers, which read pixels, and it has nothing to do with YouTube disclosure rules, which are a policy obligation rather than a file check.
Why does LinkedIn show more detail than Instagram?
LinkedIn surfaces a clickable CR badge that reveals the tool involved and who signed the credential, while Meta shows a plainer AI info entry. Both read the same underlying manifest; they differ in how much of it they choose to display.
Is Pinterest harder to get an unlabelled image onto?
It behaves differently rather than being stricter. Pinterest runs classifiers over the pixels for content that arrives without metadata, so unlike the manifest-reading platforms, cleaning the file removes only one of two inputs.
Can I remove a label from something already posted?
No. Platforms read the file at upload, and the label then belongs to that post. Cleaning the file afterwards changes nothing, and deleting and re-uploading is the only route — at the cost of the engagement the original post had.
Why do Midjourney images often arrive unlabelled?
As of early 2026 Midjourney does not embed C2PA Content Credentials, so there is no manifest for a platform to read. Tools that do embed them, such as Adobe Firefly and DALL-E 3, produce files that platforms label reliably.
Does the YouTube rule apply to still images?
The disclosure policy covers realistic altered or synthetic content in videos, Shorts and livestreams where it could mislead viewers about real events, people or places. It is about the content rather than about any metadata in the file.
If a post has no AI label, does that mean it is a real photograph?
No, and this is the weakest part of the system. A label depends on the generating tool having written metadata and the platform having read it. An image from a generator that writes nothing arrives unlabelled, so absence of a label is not evidence of camera origin.