Does OpenAI use SynthID? What OpenAI Verify checks, and what it cannot
OpenAI says yes. Its help article says images made with ChatGPT, Codex or the OpenAI API include both Content Credentials and a SynthID watermark, and its Verify page looks for them. A found signal is evidence about origin only, and a missing one rules nothing out. Our checker reads Content Credentials but not SynthID.
Does OpenAI use SynthID?
OpenAI says it does. Its help article on provenance signals says images generated with ChatGPT, Codex or the OpenAI API include Content Credentials and SynthID watermarks. It says audio from ChatGPT or the API carries a SynthID watermark. It says SynthID was developed by Google DeepMind.
| Content type | Signals OpenAI says it uses |
|---|---|
| Images | Content Credentials and SynthID |
| Audio | SynthID |
| Text | textGrain, OpenAI's own method (ChatGPT text in the EU; API opt-in) |
OpenAI also says coverage varies by product, model, export path, file type and the date the content was created. For the API, it says outputs are covered except for models planned for imminent deprecation, and it lists gpt-image-1, gpt-image-1.5 and sora-2 as examples.
Is SynthID OpenAI's watermark or Google's?
The technology is Google's: OpenAI's help article says Google DeepMind developed it. In a post dated October 7, 2026 and updated October 29, 2026, Google says anyone can check whether a file was made with AI from Google or its partners, including OpenAI, NVIDIA, Kakao and Apple.
The detector's own page, synthid.com, says it is not a general AI detector and can only find media from companies that have adopted SynthID. So, according to Google and OpenAI, two sites can read the watermark in an OpenAI image: Google's detector and OpenAI's Verify page.
What does OpenAI Verify check?
OpenAI's Verify page takes one image or audio file at a time. For images it accepts PNG, JPG and WEBP. It looks for C2PA metadata and a SynthID watermark, and it reports whether it found either or neither.
OpenAI says a detected signal indicates the content likely came from OpenAI tools. Its Verify page also says "Detected signals are reliable, and false positives are rare." That is OpenAI's claim; we have not tested it.
OpenAI's API guide describes a rule for its API about the signed record. A record counts as detected only when a trusted or valid manifest names OpenAI as its issuer and includes an AI-generation action. A record from another company, or one that is invalid, comes back as not detected.
Like every outside link from our checker, it opens another company's website, and an image you upload goes to them, not to us.
What can OpenAI Verify not tell you?
OpenAI's help article and Verify page are open about limits. In our words:
- Not a verdict on the content. A detected signal does not confirm the content is accurate, unedited, legally owned or shown in the right context.
- Not about other makers. The tool is not designed to find content from other companies' models.
- Not proof when nothing is found. OpenAI lists reasons an OpenAI file can still come back with no signal: it predates the signals, it came from an unsupported product or export path, its metadata was stripped, or its watermark was degraded.
How well does the watermark survive edits?
OpenAI says SynthID is designed to withstand common changes, such as light edits, copying and pasting, cropping or screenshots. It also says heavy cropping and file compression can make a watermark undetectable. Its Verify page words the goal as standing up to cropping, filters and lossy compression. Treat both as design aims, not promises.
For Content Credentials, OpenAI says they can be lost through everyday actions, such as taking a screenshot or converting a file. OpenAI advises checking the original exported file, uncropped and unconverted.
What did we find in ChatGPT files?
We hold four real ChatGPT files: one vendor sample and three filed on Wikimedia Commons under "Images generated by ChatGPT". Third parties uploaded the Commons files, so how they were handled before upload is unknown. We ran all four through our checker in headless Chrome on October 30, 2026. Our accuracy page lists each real file that carries a record.
| Files | Name in the certificate | Headline |
|---|---|---|
| 1 Commons PNG | OpenAI OpCo, LLC | Trusted Content Credentials declare AI use |
| 2 files: the vendor sample and 1 Commons PNG | OpenAI (certificates expired) | Content Credentials found; signer trust not established |
| 1 Commons JPEG | No record | Metadata found; source not established |
So 3 of the 4 carry a Content Credentials record. One reads as trusted and two do not. No certificate in our copies of the trust lists names OpenAI. The trusted file is trusted because its certificate chains, through an authority named Trufo, to a root on the C2PA Trust List. The other two certificates had expired. A picture we drew and signed with a self-made certificate naming "OpenAI OpCo, LLC" reads the same as those two, so that headline cannot separate them.
These counts describe four files. They do not show what OpenAI adds to every image. We do not read SynthID, so we cannot say whether any of the four carries it, and our measurements include no OpenAI Verify results.
What this means when you check an image
Open the checker and choose the file. Every result carries the line "Not checked here: Google's SynthID watermark (also used by OpenAI, NVIDIA and Kakao, and by Apple according to Google), Meta's watermark, Adobe TrustMark, visible marks, and pixel-based AI detection. The links below check some of these." The results page always links Google's SynthID Detector. It adds OpenAI Verify when a file's signer, claim generator or metadata names OpenAI, ChatGPT, DALL-E or GPT.
- "Trusted Content Credentials declare AI use" means the signed record is intact, its signer is on the current C2PA Trust List, and the record declares a generative source. It does not say the claim is true.
- "Content Credentials found; signer trust not established" means a signed record is present, but no list we load vouches for the certificate. It does not say the signer did wrong.
- "No supported provenance signals found" means we found no record, no source label in metadata and no Stable Diffusion watermark pattern. It does not mean the image is real or human-made, and it says nothing about SynthID, which we did not check.
Whichever headline you get, our limits line applies: these checks cannot determine whether an image is AI-generated. For a file you believe came from ChatGPT, OpenAI Verify is the maker's own test for its signals.
Guides on this topic
- What is SynthID? is the main page on this topic.
- Does ChatGPT add a watermark or metadata? covers what OpenAI says it adds to ChatGPT images.
- How to check an image for SynthID covers Gemini and Search.
- SynthID for video and audio covers the other media types.
- What "signer trust not established" means explains the second headline above.
Frequently asked questions
- Does ChatGPT put SynthID in every image?
- OpenAI says images from ChatGPT, Codex and the API include it, and that coverage varies by product, model, export path, file type and creation date. We cannot confirm it for any file, because we do not read SynthID.
- If OpenAI Verify finds nothing, was the image made elsewhere?
- Not necessarily. OpenAI says an OpenAI file can come back with no signal if it predates the signals, lost its metadata or had its watermark degraded. The tool also does not look for other companies' signals.
- Is OpenAI Verify the same as Google's SynthID Detector?
- No. Verify is OpenAI's page and looks for OpenAI's signals, both C2PA metadata and SynthID. Google's detector is Google's page and covers SynthID media from Google and its partners, which Google says include OpenAI.
- Does our checker send your image to OpenAI?
- No. It links to OpenAI Verify, and if you follow the link you upload the file to OpenAI yourself.