You see a photo that looks a little too perfect. The hands are fine this time. Nothing is melting. But something feels off.

Eyeballing doesn't work anymore. The models got too good. So you have to check the file, not the pixels.

Start with the file's metadata

AI generators now bake labels into the file itself. The big one is Content Credentials (built on the C2PA standard): a signed record saying which tool made the image, and sometimes the prompt that made it.

Google's tools add SynthID watermarks. OpenAI's generator added C2PA plus SynthID this year. Adobe, Microsoft, and camera makers like Sony, Leica, and Canon are on board too.

If the image carries these labels, checking is easy. You just need to read them.

The free tools that read the labels

Upload the original file to verify.contentauthenticity.org. It shows the Content Credentials: what made the image, when, and the edit history.

For Google-made images, SynthID has its own detector at synthid.com . Sign in, upload the file, and it shows which parts carry the watermark.

The Gemini app can check too. Upload the image in a new chat and ask if it's AI-generated. It answers from the embedded records when they're present.

Why the original file matters

Here's the catch: screenshots and re-uploads often strip the labels.

Save an AI image via screenshot and the metadata stays with the screen, not the file. Resize it, convert the format, or run it through a social app's compressor and the signed record usually dies. SynthID survives more abuse than C2PA, but only Google's tools can read it.

Always check the original download. If someone sends a screenshot of a photo, ask for the file they downloaded. The screenshot proves nothing either way.

A clean result means "unknown," not "real"

This is the part people get wrong. A checker that finds no AI labels is telling you one thing: this copy has no supported record.

That is not proof a camera took it. Anyone can strip metadata in seconds. The absence of a label is the absence of evidence, not evidence of absence.

Read it this way: labels present means AI, almost certainly. Labels absent means you don't know.

The honest workflow

Get the original file. Run it through a credential reader. If it says AI, believe it. If it says nothing, say you don't know.

And if you make AI images yourself, export or download the original file every time. Screenshots turn your labeled work into unlabeled mystery meat, and the new transparency rules in California and the EU expect machine-readable labels on AI content. For format conversions along the way, Jiggle X Wiggle's free converter does the job in seconds.

The eyes can't tell anymore. The file can, if you ask it the right way.