How to Use Google's SynthID Detector

You have an image in a group chat, a document from a vendor, or a video frame someone says came from a generative tool. The tempting move is to upload the copy you have and treat the result as a verdict. That is not careful enough. Google's SynthID Detector can help when SynthID is the question, but the file you choose, the checks you run first, and the words you use afterward all matter.

This page is a workflow, not a promise of certainty. Check AI Media is not Google, and this site does not decode SynthID. We can help you inspect ordinary metadata and C2PA Content Credentials in the browser. For SynthID itself, use Google's official detector and read its result narrowly.

Start with the least-edited file you can get

Before opening any portal, pause and look at the file chain. Is this the original export, or is it a screenshot of a post? Did it come from a messaging app? Was it downloaded from a social platform that recompresses media? Every extra hop can strip metadata, rewrite pixels, or detach credentials. That does not mean the file becomes useless. It does mean a later copy is weaker evidence than an earlier one.

Ask for the closest available source file. If there are several versions, keep them instead of replacing one with another. A raw camera file, an original generated export, and a platform download may each carry different clues. Name them clearly so you do not mix them up. If you need to share the file with someone else, send a copy and preserve your starting version unchanged.

Do not edit the file to make it upload more cleanly unless you have no other choice. Cropping, resizing, converting formats, or running a cleanup tool may change exactly the evidence you are trying to inspect. If the file is too large or in a format the official tool does not accept, note that limitation plainly rather than pretending the altered copy has the same evidentiary weight.

Run local checks before you upload anywhere

There is value in looking at simpler signals first. Use the free browser checks on this site to see whether the file carries ordinary metadata or signed credentials. The EXIF metadata viewer can show fields such as software names, timestamps, camera data, and editing history when those fields are still present. It may also show nothing useful, especially after a file has moved through apps that remove metadata.

Then try the C2PA Content Credentials checker. C2PA is different from SynthID. It is a signed provenance system that can say a file carries a verifiable manifest when one is present. A valid credential is a real clue. A missing credential is not proof that a file is human-made, AI-free, or untouched. Many honest files have no credentials, and many platforms still do not preserve them well.

These local checks also help with privacy. You learn what is embedded in the file before sending it to another service. If the metadata includes sensitive location, device, or workflow information, you can make a deliberate decision about what to upload and what to keep private. That is one reason a careful SynthID detector how-to should not begin with “just upload it.”

Use Google's official portal for the SynthID question

When the question is specifically whether a file carries a SynthID signal, go to Google's official SynthID Detector. Follow Google's current instructions there, including any limits on formats, file size, account access, or supported media. Those details can change, so treat the portal and Google's own documentation as the source of truth.

Upload the least-edited file you can reasonably use. If you have several versions, test the original-looking one first and keep notes about which copy was checked. Do not assume that a result from a thumbnail, screenshot, or heavily recompressed download applies equally to the original. The file you upload is the file being evaluated.

If the official tool reports a SynthID-related signal, describe it as a signal found by Google's tool. Do not stretch that into a claim about the author, the prompt, the full editing chain, or every system that may have touched the file. A watermark-style signal can support provenance, but it does not tell the entire story of a piece of media.

Read a negative result with restraint

The most common mistake is treating “nothing found” as “not AI.” Missing signals do not prove that. The file may never have carried SynthID. It may have come from a system that uses another provenance method. It may have been edited, transcoded, cropped, screenshotted, or copied through a service that changed the evidence. Or the official detector may simply not support the exact file type or history in front of you.

Use careful language in reports, emails, and moderation notes. “Google's tool did not report a SynthID signal in the tested copy” is much stronger than “this is real.” “The tested file carries a checkable signal” is safer than “this is definitely AI.” If you need the larger map of what different systems can and cannot show, read what can be detected in AI media. If your concern is whether edits can damage signals, the guide to why AI watermarks survive some changes gives the right context.

Write down the evidence trail

A good check ends with notes, not just a screenshot of a result. Record the filename, where the file came from, whether it was the original or a copy, which local checks you ran, and what the official portal reported. If you changed the file to make it upload, say so. If the upload failed, say that too.

That habit protects everyone involved. It keeps you from overclaiming. It lets another reviewer repeat the check later. It also makes room for uncertainty, which is part of honest provenance work. The goal is not to sound certain when the evidence is incomplete. The goal is to say exactly what the file carried, exactly what tool read it, and exactly where the evidence stops.