SynthID Detector explained: what Google’s watermark checker can and cannot tell you
On October 7, 2026 Google opened SynthID Detector to everyone. The portal at synthid.com checks an uploaded image, video or audio file for SynthID, the invisible watermark that Google and its partners OpenAI, NVIDIA and Kakao embed in AI-generated media, with Apple listed as coming soon. It is a strong check for that one signal and deliberately not a general AI detector. This page collects the facts, explains how to read a result, and shows where it fits next to metadata checks like the one AICheck365 runs in your browser.
What is SynthID Detector?
SynthID Detector is Google DeepMind’s verification portal for SynthID watermarks. You sign in, upload a file, and it reports “watermark detected” or “watermark not detected”; for video and audio it also marks which segments carry the watermark. A detection means the file was likely generated or edited by a SynthID-enabled model. A miss proves nothing about human authorship: the file may come from a model that does not use SynthID, or the watermark may have been degraded by heavy editing.
Key facts as of October 2026
| Operator | Google DeepMind. Announced by Pushmeet Kohli, VP of Science and Strategic Initiatives. |
|---|---|
| Public since | October 7, 2026, worldwide, interface in English only. An earlier version had been limited to journalists, researchers and other waitlisted testers since May 2025. |
| Address | synthid.com |
| Companies covered | Google, OpenAI, NVIDIA and Kakao. Apple is listed as “soon”. Media from tools that do not embed SynthID cannot be identified. |
| Images | JPG, JPEG, PNG, BMP, WEBP, AVIF, HEIC, HEIF, TIFF, TIF, GIF. Result: watermark present or not. |
| Video | MP4, MOV, WEBM. Result: which segments carry the watermark. |
| Audio | WAV, MP3, OGG, FLAC, AAC, M4A. Result: which segments carry the watermark. |
| Text | Not accepted by the portal, although SynthID also has a text watermark. |
| Account | Sign-in required; third-party sign-in such as Apple is supported. Daily usage quotas apply. |
| Privacy | Uploads are processed in real time and deleted once the result is returned. A digital signature of the file is kept for 24 hours to enforce quotas. |
| Scale | Google says SynthID has marked more than 180 billion images and videos and 240,000 years of audio, and that SynthID checks in Search, Chrome and the Gemini app handle over 1 million requests a day. |
How to check a file with SynthID Detector
1. Find the best copy
Use the original download from the generator, not a screenshot, chat-app copy or re-encoded repost. Google asks for the highest-quality file you have because heavy compression and editing weaken detection.
2. Sign in at synthid.com
The detector needs an account so it can apply daily quotas. Accept the additional terms only if you are fine with them.
3. Upload one file
Pick an image, video or audio file in a supported format. Video and audio results are segment-based, so long files can show watermarked and clean parts side by side.
4. Record the result with context
Note the date, the file you tested and where it came from. A watermark result is one data point; Google itself recommends collecting several before deciding.
How to read a SynthID Detector result
The portal returns a binary answer. The wording below follows the synthid.com FAQ.
| Watermark detected | A SynthID watermark is present. The content was likely generated or edited by a SynthID-enabled model from Google, OpenAI, NVIDIA or Kakao. It does not tell you which product, prompt or person made it, or whether only part of an image was edited. |
|---|---|
| Watermark not detected | No confident signal. The file was probably not made by a SynthID-enabled model, but AI generation is not ruled out: other generators such as Midjourney or Stable Diffusion do not use SynthID, older content predates partner adoption, and heavy modification can destroy the signal. |
| False positive | Google says that in very rare cases SynthID can trigger on content that was never watermarked. Treat a single positive as strong evidence, not as a verdict. |
SynthID vs metadata checks vs general AI detectors
These methods answer different questions. None of them covers every file, which is why they work best together.
| Method | What it reads | Strength | Blind spot |
|---|---|---|---|
| SynthID Detector (synthid.com) | Invisible watermark woven into pixels, video frames or audio frequencies | Survives filters, colour changes, lossy compression and metadata stripping | Only covers SynthID adopters; file must be uploaded to Google; no model or prompt details |
| C2PA and metadata (AICheck365) | Content Credentials, digitalSourceType, EXIF, XMP, PNG parameters, MP4 fields | Names the tool, can carry a signature, runs locally without uploading | Lost after screenshots, chat-app compression or re-export |
| General AI classifiers | Statistical artifacts in pixels or audio | Works when no signal was embedded | Probabilistic, weaker on new models and recompressed files, hard to use as proof |
| Google Cloud AI Content Detection API | Pixel-level artifacts, noise and spectral patterns (images only) | Programmatic, covers Google and third-party models | Private Preview by application; a classifier, not a SynthID watermark lookup |
Recommended workflow: metadata first, watermark second
SynthID and file metadata fail in opposite situations. Metadata disappears when a file is screenshotted or recompressed, while the watermark is designed to survive exactly that. Metadata can name the tool and carry a signature; the watermark cannot. Use both:
- Run the original file through the AICheck365 detector. It reads C2PA, EXIF, XMP, PNG text chunks and MP4 metadata locally, so nothing leaves your browser.
- If the scan finds a Google, OpenAI, NVIDIA or Kakao signal, upload the same file to SynthID Detector to confirm the watermark is still present in this copy.
- If the scan finds nothing because the file was shared or screenshotted, SynthID Detector is the next best check for media that may have come from those companies.
- If both come back empty, the file may still be AI-generated by a tool that writes neither metadata nor SynthID. Look for the earliest source and compare other copies.
API and integration status
synthid.com has no public API. Its additional terms forbid submitting requests programmatically, scraping, or using the frontend as an automated API, and they forbid creating extra accounts to get around quotas. Any integration has to send people to the portal rather than call it behind the scenes.
For businesses, Google lists an AI Content Detection API on its Gemini Enterprise Agent Platform. According to its documentation it is in Private Preview, accepts JPEG, PNG and WebP images, and uses models that analyse pixel artifacts, noise and spectral anomalies, which makes it a classifier rather than a SynthID watermark check. Access requires an application.
SynthID Text, the text variant, has been open-sourced by Google, but the Detector portal does not accept text uploads.
Where AICheck365 fits
AICheck365 does not read SynthID. The watermark’s decoder is proprietary, and we do not upload files. What we read is the provenance data that travels in the file itself, including C2PA manifests that sometimes state SynthID was applied. That statement is a claim recorded by the generator, not proof the watermark survives in your copy. When our scan is empty or points to a SynthID partner, the result page now links to SynthID Detector as the next step.
Check the file locally first, then decide whether a watermark check is worth an upload.
Frequently asked questions
Is SynthID Detector free?
Google does not list a price. You need to sign in, and daily usage quotas apply.
Can SynthID Detector tell if an image is from ChatGPT?
Only if the image carries a SynthID watermark. OpenAI is listed as a SynthID adopter, so images from OpenAI products that embed it can be detected; images made before adoption, or by other tools, will show as not detected.
Can it detect Midjourney or Stable Diffusion images?
No. Those tools are not SynthID partners. Check the file’s metadata instead, for example PNG generation parameters or C2PA, with a metadata reader such as AICheck365.
Does “watermark not detected” mean a human made it?
No. It only means no confident SynthID signal was found. The file may be from a non-partner model, from before adoption, or heavily modified.
Does SynthID survive screenshots and compression?
It is designed to survive filters, colour changes and lossy compression, which usually strip metadata. Google warns that it is not infallible and that many combined transformations can remove it.
Does Google keep the files I upload?
According to the synthid.com privacy notice, files are deleted once the result is returned. A digital signature of the file is kept for 24 hours to enforce quotas and is then deleted.
Can SynthID Detector check text?
Not at the moment. The portal accepts images, video and audio. SynthID also has a text watermark, but text checking is not offered on synthid.com.
Is there a SynthID Detector API?
Not for the portal, and its terms forbid automating it. Google’s separate AI Content Detection API for businesses is in Private Preview and is an image classifier rather than a watermark lookup.
Sources
- Google: We’re making it easier to identify AI-generated content globally (Oct 7, 2026) — Public launch, partners, scale figures.
- SynthID Detector — Supported formats, FAQ, additional terms and privacy notice.
- Google: SynthID Detector early access (May 20, 2025) — Original announcement for early testers.
- Google Cloud: Detect AI-generated images — AI Content Detection API, Private Preview.
- Google DeepMind: SynthID — Overview of watermarking for image, audio, text and video.