How to check AI image watermarks with SynthID
Use Google SynthID Detector to check supported AI watermarks. See one recorded image result, interpret no detection and review ecommerce media before publishing.
Start with a narrower question than “Is this image AI?”
SynthID Detector checks for supported AI watermarks. It does not establish the authorship of every image on the internet. Google announced global English-language access on 7 October 2026 for image, video and audio checks, naming Google, OpenAI, NVIDIA and Kakao, with Apple upcoming. Those names do not establish coverage of every provider model or historical export. Use the tool to gather one source signal, alongside the original asset and its creation record.
What our one-image check actually showed
On 8 October 2026, we submitted one Google Gemini Image showcase JPEG through the SynthID web interface. The first result said “SynthID was detected” and “This media was made or edited with Google AI”. The interface also warned that the media may have been edited further. The 172,827-byte input, source URL, SHA-256 hash, first result and screenshot are linked in Research. We did not generate the image ourselves; its exact generation date and original private export were unavailable. The source CDN may have resized or compressed it.
How to check one permitted image
Keep a copy of the highest-quality permitted source file and record where it came from. If it is an AI export, retain the creating product, displayed model/version and date when available. Save a file hash so later reviewers know which bytes were tested. Open the official portal, confirm access and current upload restrictions, then submit the file through the intended interface. Preserve the exact result or error before editing the asset. uAgentKit links to the detector and records evidence; it does not run an embedded detector or a batch verification API.
Interpret the result without inventing certainty
The rows below are explanatory categories, not a quotation of every possible interface label and not additional test results.
| Observation | Reasonable interpretation | What it does not establish |
|---|---|---|
| A supported watermark is reported | Retain the reported provider and any displayed file or segment scope. | Commercial rights, product fidelity, a complete edit history or marketplace approval. |
| No supported watermark is reported | The submission did not establish a supported watermark in this check. | Human authorship, originality, authenticity or absence of all AI editing. |
| The file is rejected or processing fails | No usable detection result was obtained; retain the exact error. | A positive or negative watermark finding. |
Use it in an ecommerce media-review workflow
Before publishing a supplier image or campaign clip, retain the asset source and permissions. Record the watermark result, then compare the asset with the actual SKU: variant, packaging text, materials, proportions and claims. Check licensing and the destination platform’s current rules separately. A detected watermark does not automatically prohibit commercial use, and no detection does not grant permission. Assign a person to approve publication using the whole record.
Free access, formats and limits: what remains unverified
Public availability is not proof of an unlimited free plan. We did not measure a per-file charge, quota, upload-size ceiling or typical processing time. The inspected file picker listed image, audio and video extensions, but a picker list is not evidence that every format works. Only the named JPEG was submitted here. The broader SynthID technology includes text watermarking; that fact alone does not establish a text-upload feature in this public portal.
SynthID Detector versus a general AI-content detector
Google describes SynthID as adding an imperceptible watermark during content creation and detecting supported watermarks later. A general claim that software can identify any AI image is broader than this documented scope. Do not treat no detection as a finding that a file is human-made. To compare detectors, define the same permitted files, provenance and decision first, and retain each result without turning a tiny sample into a ranking.
A quick test is a starting point
This is one convenience sample, not an accuracy benchmark. A permitted camera-original control, independently documented model exports, and separate video and audio checks would add evidence, but those tests were not performed in this record. Preserve failures and unknowns, disclose account context, and report only observed outputs. The linked record remains separate from the product’s broader predefined test plan.
Continue your comparison
Google SynthID Detector: features, access and limits →
Original one-image detection record →
Related products and tasks
Google SynthID Detector · Explore the related task →
Sources supporting this guide
- Google announces public SynthID Detector access
Google · blog.google · Read · 2026-10-08
- Google DeepMind explains SynthID
Google · deepmind.google · Read · 2026-10-08
Research method & original evidence · Who operates uAgentKit · Commercial disclosure