A screenshot is not a reliable proof of SynthID removal. It makes a new image of what an app displayed. That can discard the original file’s attached records and change its pixels, but neither change tells you whether a watermark is still detectable.
The right question is narrower: does the exact screenshot trigger the relevant verifier, and what detail did the screenshot lose?
What a screenshot actually copies
A screenshot normally captures a rendered view, not the original image file. If a large image is fitted into a small window, the screenshot records that smaller rendition. It may also include interface elements, a background, a cursor or a visible watermark.
The screen view can depend on zoom, display scaling, color management and the application. Saving the screenshot adds another file-format decision. Sharing it can add more resizing or compression.
Illustrative chain. Every arrow can introduce a change unrelated to watermark removal.
Source
What this means
Zoom, fit-to-window sizing, color management and overlays can change the rendered view.
Capture
What this means
The screenshot may contain the same visible content at fewer pixels, plus app borders or controls.
Share
What this means
That copy needs its own check. A screenshot operation alone supplies no detector verdict.
That chain explains why “I took a screenshot” is not a complete experimental description. Two people can capture the same source image at different resolutions and produce different files.
Which information changes?
| Property | What a screenshot may do | What you still need to check |
|---|---|---|
| Original EXIF or XMP | Usually does not copy those records as source metadata | New screenshot metadata and the actual exported file |
| Visible logo | Captures it if it remains on screen | Whether the logo is still visible |
| Image dimensions | Uses the captured display area | Whether the result has enough detail |
| Pixel-level watermark | Produces a changed rendition of the content | The exact screenshot’s verification result |
| Visual AI classification | Supplies a different file to the model | That classifier’s result |
A screenshot can contain sensitive information even without the original GPS tags. A street sign, address, account name or browser tab title can remain visible. That is why metadata privacy and visual privacy need separate checks.
A concrete quality example
Imagine a 3,000-pixel-wide product image shown at 1,000 pixels wide in a browser. A screenshot of that view cannot recover the full original spatial detail merely because you later save it as a large PNG. Enlarging the screenshot estimates more samples from the smaller capture; it does not recreate the missing fine lettering.
This is an illustrative example, not a measured SynthID test. It shows a cost that can be assessed independently of a detector. For a product catalog, loss of a label or fabric texture can make the screenshot unsuitable even if a score changes.
Inspect at the final use size first. Then compare close details at matching zoom. Comparing a small screenshot to a much larger original in separate windows can make differences look smaller or larger than they are.
How to test the screenshot claim fairly
Keep the original, then create a controlled display and capture. Record the displayed size, screenshot dimensions, crop and export format. Verify the original before making claims about the screenshot.
Add an ordinary resized export at the same dimensions as the screenshot. If both behave similarly, resizing or encoding may explain the result more simply than the screen-capture step. If they differ, you still need to examine other changes such as color handling or borders.
Avoid testing ten screenshots and reporting only the one favorable response. That selects an outcome without showing how often the workflow failed. For a personal check, keep the exact successful file. For a study, keep the whole set of attempts and the selection rule.
What our existing experiments contribute
Deprint’s frozen experiments do not establish a screenshot removal rate. They do show why export controls matter. Size-matched controls for the desk and bakery did not reproduce a broad detector bypass: the documented desk Sightengine result remained 99%, and the bakery Hive result remained 99.9%.
Those were resizing controls, not screenshot trials. Their relevance is methodological: changes in resolution should be isolated instead of credited automatically to another operation. See the recorded methods for file and detector distinctions.
A better decision for everyday use
If your goal is a clean sharing copy with fewer attached records, a direct metadata-aware export can preserve more quality than a screenshot. If your goal is a watermark check, verify the delivered file. If your goal is a smaller attachment, compare compression options at the required dimensions.
A screenshot may be convenient for showing something on your screen. Convenience is different from demonstrated watermark removal. Keep the original so that a quick capture does not become the only version you have.