How We Check Files

Someone sends you an image and asks, “Can you tell if this came from AI?” That sounds like a simple yes-or-no request, but it is not how trustworthy checking works. A file can carry useful clues, lose them during sharing, or never include them in the first place. Our method starts there: read the evidence that is actually inside or attached to the file, then say plainly what that evidence can and cannot support.

We read signals, not intentions

Check AI Media looks for checkable signals a file may carry. That can include ordinary metadata fields, signed Content Credentials, and visible hints that point a person toward an official watermark checker. Those are different kinds of evidence. A camera setting in metadata is not the same as a signed provenance record. A label in an image is not the same as a private detector result. We keep those categories separate because mixing them together makes the answer sound stronger than it is.

When a file carries AI credentials, we can say that the file carries them. That is a useful finding. It may tell you that a creator, tool, or publishing system attached provenance information to the file. It does not let us reconstruct every edit, every export, or every platform that touched the file unless those details are present in the record. The honest answer is narrower than people often want, but narrower is better than misleading.

When no relevant signal is found, the result is also narrow. It means this check did not find the kind of signal it can read. It does not mean the file is human-made. It does not mean AI was not involved. Metadata can be stripped by messaging apps, websites, screenshots, resizing tools, and normal export settings. Some AI tools may not attach credentials at all. Some human-made files may contain little or no metadata. Absence is not a clean verdict.

What happens in the browser

The current browser tools are designed so the file stays on your device while the check runs. For a person using the EXIF metadata viewer or the Content Credentials checker, that matters. You may be checking a private draft, a client image, a family photo, or a file you do not have permission to share more widely. The tool should not ask for more trust than the job requires.

That privacy choice also shapes the way we describe results. We do not send the file away for a hidden second opinion. We do not compare it against a private database. We do not make a behind-the-scenes judgment about authorship. The page reads the available signals in the file and explains them in ordinary language. If a future feature requires upload, such as a server-side video check, it should be labelled before you choose a file so you can decide whether that tradeoff is acceptable.

Browser-based checking is not magic. Some file types expose more information than others. Some exports are tidy; others are messy. A screenshot of a credentialed image may no longer contain the credential. A reposted copy may have been recompressed. A file can be edited in a way that preserves some information and removes other information. The method is practical, not theatrical: inspect what is available, avoid filling gaps with guesses, and tell you when the gap matters.

Why we are strict about SynthID

SynthID needs special care because people search for it as if every website can read it. We do not claim to detect SynthID ourselves. As of this writing, Google points people to its official SynthID Detector for that purpose, and Google has not published a public detection API that would let an independent site reliably read SynthID in the same way. If you specifically need a SynthID answer, use Google’s official route rather than a random “detector” page that promises more than it can show.

We may discuss SynthID on this site to reduce confusion, but that is not the same as performing the check. The distinction matters. A visible mark, a platform label, or a file name that mentions AI may be useful context for a person reviewing a file, but it is not a verified SynthID read. If you want a deeper explanation of that boundary, the SynthID detector guide explains what an independent site should and should not claim.

We also do not help remove watermarks or credentials. A provenance signal exists to preserve context. Removing it can mislead viewers, clients, editors, or audiences. Sites that advertise watermark removal often blur the line between legitimate format conversion and concealment. Our position is simple: check signals, preserve context, and do not use a tool whose main promise is to make provenance harder to see.

How to use a result responsibly

A useful result is one part of a review, not the whole review. If a file carries signed credentials, keep the original copy and share the result with the people who need it. If the check finds ordinary metadata, treat it as context rather than a final answer. If the check finds nothing, do not turn that into a claim of human authorship. The safest wording is usually precise: “This file carried no readable credentials in this check,” or “This file carries Content Credentials that say AI was involved.”

For newsroom, classroom, marketplace, or client work, keep the chain of custody as clean as you can. Ask for the original export when possible. Avoid screenshots if you need provenance. Save a copy before uploading to social platforms. Write down where the file came from and what tool you used to check it. Those habits are not glamorous, but they make later conversations less slippery.

If you are unsure which check fits your situation, start with the general what can be detected page. It separates credentials, metadata, visible labels, and official watermark tools. That is more useful than pretending there is one master detector that can settle every file.

The boundary we will not cross

The line is this: we can report checkable signals when they are present, and we can explain why missing signals do not prove anything about human authorship. We will not call a file “not AI” because a browser check came back empty. We will not imply affiliation with Google. We will not dress uncertainty up as a percentage. Credibility through honesty means leaving some questions open when the file does not contain enough evidence to close them.

That may feel less satisfying than a dramatic verdict, but it is the only kind of answer worth relying on. If the file carries evidence, use it. If it does not, say that carefully and keep looking for better context.