Use a compatible watermark verifier and check the exact image file you care about. Record the result as detected, not detected or unavailable. A general AI image classifier, a metadata inspector and a conversational guess about visual style are not interchangeable with watermark verification.
If you are comparing an original with a processed copy, check both. Without a positive starting result, a negative output cannot demonstrate that processing removed anything.
Choose the check that matches the question
Google documents verification through its supported interfaces, including Gemini and the SynthID Detector. Interface availability and supported media can change. Start from Google’s current product or help page rather than an old screenshot of a menu.
When using a conversational interface, ask specifically whether a SynthID watermark was detected in the attached image. Read the answer closely. A comment that the picture “looks AI-generated” is a visual assessment, not necessarily a report from the watermark tool.
| Question | Appropriate evidence | Inadequate substitute |
|---|---|---|
| Is a supported watermark detected? | A completed compatible verification response | An empty EXIF report |
| Did processing change detectability? | Original and output checked with a positive baseline | Only a negative output |
| Does a visual classifier flag the output? | That classifier’s result for the output | A SynthID-negative response |
| Is the picture a camera original? | Reliable source history and relevant supporting records | One negative detector result |
Keep the result states separate
Illustrative verification log. “Unavailable” must not be counted as “not detected.”
Detected
What this means
Record the tool, date and exact file.
Not detected
What this means
Keep the wording narrow. This does not establish the origin of every part of the image.
No usable result
What this means
Retrying later can produce a result, but the missing check contributes no success evidence.
An error belongs in your log. If a tool times out, rejects the format or reaches a quota, you have learned about the attempt, not the image’s watermark. Counting that event as “passed” would make a broken detector look like successful processing.
Ambiguous wording also needs care. “I cannot determine that” does not mean “I checked and found no watermark.” Save the exact text instead of forcing every reply into a yes/no column.
A repeatable before-and-after workflow
- Save the original download and make it read-only or keep a separate untouched copy.
- Check that original in the chosen verification interface. Record the date and response.
- If you are testing an editor, process a copy with documented settings.
- Download the actual output. Do not substitute the editor’s thumbnail or a screenshot of its preview.
- Verify the output through the same interface and keep the response.
- Inspect visual quality separately, including text, faces, color and dimensions.
For more rigorous work, record hashes, file sizes, dimensions and format. This makes it possible to distinguish “the same scene” from “the same file.” A JPEG and a PNG of the same portrait are related images, but not the same submitted bytes.
Use a control when metadata is part of the question
Suppose you want to know whether a metadata cleanup changes the outcome. Compare the original with a copy that retains the original decoded pixels while removing the relevant records. Check that pixel preservation with a consistent decoding procedure; do not assume it from the tool’s label.
If both remain positive, you have evidence that deleting those records was insufficient for that image and verifier. If the result changes, investigate whether the export changed pixels, dimensions, orientation or color handling as well. A bundled export operation can make the cause difficult to identify.
In Deprint’s archived desk and bakery experiments, original-pixel controls without attached metadata remained positive in Gemini. Those controls helped distinguish the later pixel treatments from metadata removal. The methods retain the exact scope and limitations.
Why the interface and date belong in the result
A verification service is part of the measurement. Its upload handling, supported formats, model and response format can evolve. Even when the service name stays the same, a result collected months later may not be directly comparable.
The frozen Deprint observations used Gemini’s web interface. Its internal handling of uploads was not visible. The archive does not turn those responses into raw watermark scores or standalone portal measurements. Describing the interface makes that boundary understandable.
How to summarize the outcome
Prefer a statement such as: “For this output, the named verifier reported no SynthID on the recorded date; the original control was positive.” That is more useful than “the image is undetectable.”
If you also checked visual classifiers, list them separately. Include negative and positive results, and label missing checks. A single row per exact file helps readers see which observations belong together.
This guide provides a verification procedure, not a guarantee that an image will receive a particular verdict. The next useful step is to design the comparison around the claim you actually want to test.