How to Check Your Image for AI Watermarks and Tags Before You Post It

There is one outcome in this whole topic you cannot undo: publishing a file and *then* discovering what it was carrying. The label appears, the metadata is already in someone's cache, and the location tag you never knew was there has already travelled.

The fix costs three minutes and happens before you hit post. Here is exactly how to audit your own image.

This is a spoke of our AI image watermarks pillar guide.

What you can and cannot check

Four of six rows are fully knowable. That is more control than most people assume they have.

Step 1: look at the picture

Obvious, and routinely skipped. Open the image at full size and check all four corners and both edge bands for a badge, logo, coloured strip or text. Marks are frequently small, low-contrast, and easy to miss in a thumbnail.

If you find one, it is geometry: crop it, or patch it from surrounding pixels.

Step 2: read the file, not the picture

This is where the informative data lives. Drop the image into the metadata tool and it enumerates what is actually present:

• C2PA manifest — a signed record of how the file was produced, sometimes including edit history
• IPTC digital source type — an explicit synthetic-origin declaration
• XMP and EXIF generator tags — software and model names, versions, processing history
• PNG text chunks — where generation parameters often end up, prompts included
• ICC profile and camera EXIF — and on phone photos, GPS coordinates and capture time

The full anatomy of these fields is in our AI image metadata guide.

Step 3: read the before-and-after report

The report does two things. It lists every signal found and confirms what removal cleared — so the metadata layer is verified rather than assumed. And it gives a heuristic detection-likelihood estimate for the pixels, computed locally from characteristics associated with generated imagery: texture smoothness, noise structure, frequency regularity and similar tells.

Be clear about what that score is. It is a local estimate, useful for comparing before against after. It is not a verdict from any platform's detector, and nobody outside a platform can give you one. Our guide to reading a detection report walks it through field by field.

Step 4: the layer you cannot verify

Invisible pixel watermarking cannot be independently confirmed, because detection requires the issuing provider's own model. That gap is deliberate: a public verifier would let anyone test edits repeatedly until detection failed, which is precisely the attack the scheme exists to resist.

So treat it as an unknown that may be present, handle the layers you can, and distrust any tool claiming to verify or guarantee removal here. Our SynthID guide explains the asymmetry in full.

Step 5: re-check the output

Run the processed file through the report again. Ten seconds, and it converts "the container should be empty" into "the container is empty". Verification is the entire reason a before-and-after view exists.

The three-minute checklist

1. 1Zoom in and check every corner and edge for a visible mark.
2. 2Inspect the metadata and read the full list of what is present.
3. 3Remove the badge if there is one; strip the container either way.
4. 4Read the detection-likelihood estimate as a heuristic, not a verdict.
5. 5Re-check the output before publishing.

Read next: how platforms turn these signals into a label on your post →