No public SynthID detector exists for arbitrary images. Google runs the decoder, and it lets the public reach it only indirectly: by asking the Gemini app a question. OpenAI has its own check for images it made. Gemini gives you no score, and neither check tells you anything useful about a picture from a third company.
So a site offering to “check if an image has SynthID” is doing one of two things. It is passing your image to one of the checks below, or it is guessing from something else in the file. We do the second. Our label says so.
The checks that exist, as of September 2026
| Check | Who can use it | What it looks for | What the answer is |
|---|---|---|---|
| Gemini app: upload the image and ask “Was this image created or edited by Google AI?” | Anyone with a Google account | Google’s SynthID watermark, plus Gemini’s own reasoning about the picture | A written answer from a chat assistant. No score, no confidence figure. |
| SynthID Detector portal | Journalists and media professionals Google has let in; others can join an early tester waitlist | SynthID in Google-made media | We haven’t had access, so we can’t describe what it returns |
| OpenAI Verify at openai.com/research/verify | Anyone | A SynthID watermark from OpenAI, or a trusted C2PA manifest from OpenAI | Whether the image came from OpenAI. OpenAI says errors happen but are rare. |
Google’s November 2025 announcement describes the Gemini route precisely: Gemini will “check for the SynthID watermark and use its own reasoning” to answer. Those are two different inputs. You get one reply, and nothing in it tells you which input produced it.
Reading a Gemini answer
Treat it as a classifier’s judgment. A “yes” means the mark was found and, in the words of Gemini’s help page, that “all or part of the image” was created or edited by Google’s AI. The page is careful to say “or part”. By that wording, a real photo with one object added by Google’s editing tools should come back positive too. A “no” is weaker. The same help page says Gemini can currently recognise only content made by Google AI tools, so a no on a Midjourney picture says nothing about Midjourney. It also lists unclear results such as “Not enough details to watermark” and “Likely too small an edit”. Screenshots are accepted, as long as they’re cropped tightly around one image.
Nobody outside Google knows how often either answer is wrong.
Google publishes no false-positive rate for the Gemini check, no threshold, and no way to see the decoder’s raw output. Our own watermark tests lean on this check, and we still can’t tell you how much of any reply comes from the watermark and how much from the model looking at the picture.
For OpenAI images the question is simpler. Since 20 May 2026, images from ChatGPT, Codex and the API carry both SynthID and C2PA, and OpenAI’s Verify page looks for either. If the C2PA manifest is still in the file, that alone can answer; if it has been stripped, the watermark is what’s left to find.
What WipeTheAI shows instead
When you add an image to our tool, the SynthID line never says “detected” or “not detected”. It can’t. We don’t have Google’s decoder, and pretending otherwise would be the exact thing we criticise above. It reads the file’s own evidence instead and gives one of three labels.
- Likely when the file names a Google AI product (Gemini, Imagen, Vertex AI, SynthID or Nano Banana), or names Google as the Content Credentials generator alongside an AI-source marker.
- Possible when the file says some other AI tool made it, and also when it says nothing at all, because a file with no metadata can still carry a watermark.
- After a wipe that changes pixels (either mode), a note that the pixels were disrupted and that no public detector can confirm the mark is gone.
One rule exists to prevent a false alarm: the word “Google” alone doesn’t count. Photos from Pixel phones carry Make: Google in their EXIF data, and calling a holiday snap “likely SynthID” because of the camera brand would be wrong. Every label comes with the same line: no public detector can confirm it.
If you need a real answer about a Google image, ask Gemini, and read what it says as one opinion with unknown error rates. For what the literature and our own tests say about disrupting the mark, see the notes on SynthID removal; for the Content Credentials side of the same file, checking an image for Content Credentials covers the tools that read them.