AIGC Labels, Explained in Plain Language

A colleague forwards an image with a one-line note: “It says AIGC, so that settles it.” Maybe. Maybe not. “AIGC” is short for AI-generated content, and an “AIGC label” is any mark, tag, or record attached to a file to say that AI was involved. That can mean a sentence printed across the picture, a hidden field inside the file, or a signed record that travels with it. Each one tells you something different, and each survives copy-paste differently.

This is why “does it have an AIGC label?” is rarely the clean yes-or-no question people expect. A screenshot can keep a visible label and lose everything hidden. A messaging app can strip a metadata tag on the way through. A creator can type a label into the caption while never writing the structured record a checker looks for. If you want the label to mean something in a real decision, it helps to know which system wrote it and who can read it back.

The label systems you will actually meet

Most of the “AIGC labels” in circulation come from a small set of systems. They are not versions of the same thing, even when people use one word for all of them.

China’s AIGC label. Under China’s labeling rules, which took effect on September 1, 2025, AI content is expected to carry both a visible mark and a structured metadata label. The metadata block is the part a file checker can read. It uses fields whose names include “AIGC” and can record whether the content is AI-generated, may be, or is suspected to be, along with a service provider and traceability IDs. You can inspect it with theChina AIGC label detector.

IPTC Digital Source Type. The IPTC, which maintains photo-metadata standards used by newsrooms and photo tools, defines a field called DigitalSourceType. Its values describe where an image came from at a controlled level of detail: captured from real life, edited by a person, composited, or created with generative AI. The clearest generative value, trainedAlgorithmicMedia, is a declaration, not a verified fact. TheDigital Source Type detector reads it.

C2PA Content Credentials. C2PA is a different idea: a signed manifest attached to a file that can describe creation and editing actions and can be checked for integrity. It is not a single word like “AIGC” but a record that can be independently verified while it remains attached and valid. TheContent Credentials checker looks for one.

Visible disclosure marks. The oldest kind of label is the one you can see: a caption, a badge overlaid on a platform, a “generated with AI” line under a post. Visible marks are useful for a human reader, but they are easy to crop out, retype, or add by hand, so they are weak evidence about how a specific file was made.

Visible and hidden labels are not the same thing

A person reading a page sees a visible label. A checker reading a file sees a hidden one. The two can disagree, and the disagreement is often mundane. Someone may add a visible “AI” note to a photo that carries no metadata at all. A platform may put its own visible badge on a clip while wiping the producer’s embedded record. A screenshot may inherit the look of a labeled image without any of its file-level information.

That is why a visible label answers a different question from a hidden one. The visible mark tells you what someone wanted the audience to know at the moment of publishing. The hidden label tells you what a specific copy of the file declares about its own production. When the stakes are high, you want both, and you want to keep them separate in how you describe the result.

How to check which label a file carries

Start with the check that matches the label you suspect. If the file may have come from a Chinese AI service, look for the structured AIGC metadata. If you are handling a photo-press workflow, look for the IPTC source type. If you want a signed record, look for Content Credentials. Ordinary fields such as software names, camera details, and dates can add context, so it is worth opening theEXIF metadata viewer as well.

Run these checks on the file itself in your browser before a platform rewrites it. The longer a file travels, the more its labels can change. If you need the whole map of signals first, readwhat can be detected in AI media before choosing a checker.

What a found label does and does not prove

A found label proves that the file carries a particular declaration. For China’s structured label, that means the fields are present and readable. For the IPTC source type, it means the file says, in a standard vocabulary, how the image began. For C2PA, it means a manifest is attached and may still verify. Those are real, quotable facts about the copy in front of you.

What a label does not do is authenticate itself simply because it exists. Ordinary metadata and XMP fields can be edited or copied. A visible mark can be added to any image. Even a signed credential should be read for what it claims and whether the signature still validates, not treated as a full history of the file. The honest sentence names the system: “This file’s metadata declares it was created with generative AI,” not “This file is confirmed AI.”

When there is no label at all

A missing label is the most common result and the easiest to misread. It does not mean the content is human-made. The tool that created it may never have written a label, or the label may be gone because the file was exported, compressed, reposted, or captured as a screenshot. Many everyday workflows erase these fields without anyone intending to.

So treat an empty result as a statement about this copy, not about the original. Ask for the earliest version you can get, compare the files you have, and keep the one that carries the strongest record. When a label is present, quote it precisely. When it is absent, say so plainly and stop there instead of filling the gap with a guess.