Which AI image generators add Content Credentials, metadata or a watermark?
Several makers say their image tools add a signed Content Credentials record, an invisible watermark, or both. OpenAI says ChatGPT, Codex and API images carry both. Google says Gemini API images carry SynthID. Adobe, Microsoft and Black Forest Labs describe Content Credentials. Midjourney describes a hidden ID tag. OpenAI and Midjourney say edits and sharing can remove these signals.
What does each maker say it adds?
The table lists the makers whose own pages we could open and read. It describes what each maker says about its output, as of October 30, 2026. Coverage depends on the product, the model and how a file was exported, so read the notes in each row.
| Generator or service | What the maker says it adds | Source |
|---|---|---|
| ChatGPT, Codex and the OpenAI API | Content Credentials and SynthID watermarks. OpenAI says coverage may vary by product, model, export path, file type and creation date. It says API output is covered except for models planned for imminent deprecation, including gpt-image-1 and gpt-image-1.5. | OpenAI help article, read October 30, 2026 |
| Gemini API (Nano Banana models) | A SynthID watermark on generated images. | Gemini API image page, read October 30, 2026 |
| Google Cloud image models | Content Credentials, signed by Google LLC, on images or videos made with the listed models. The page lists models such as gemini-2.5-flash-image and gemini-3-pro-image. | Google Cloud page, last updated October 28, 2026 |
| Adobe Firefly | Content Credentials, applied automatically where every pixel is generated with Firefly, Adobe says. The record names Adobe Inc. as issuer and may also be stored in Adobe's public cloud. | Adobe help page, last updated August 25, 2026 |
| Microsoft Azure OpenAI | Content Credentials on all AI-generated images from Azure OpenAI, with a manifest that carries a description and a software agent name. | Microsoft Learn page, dated January 30, 2026 |
| Image Creator in Paint | A C2PA manifest, which is the record behind Content Credentials. | Microsoft support page, read October 30, 2026 |
| Black Forest Labs (FLUX) | Content Credentials on output from its API, its Developer Playground and hosted integrations it controls. For FLUX.1 [dev] and [schnell] run elsewhere, it says it cannot ensure they are attached. | Black Forest Labs help article, read October 30, 2026 |
| Midjourney | A small ID tag hidden inside the file of every image it delivers. The page does not call it Content Credentials. | Midjourney documentation, undated, read October 30, 2026 |
Two notes on the table. The Google Cloud page covers Google's cloud platform, and the pages we read do not say what each consumer app adds. The Microsoft Learn page carries a "classic" label and describes the DALL-E and GPT-image-1 series.
What about generators that are not in the table?
For Ideogram, Leonardo, Grok and Seedream, our searches on October 27 and October 30, 2026 did not turn up a page on the maker's own site that describes Content Credentials or an invisible watermark. That is a limit of our search. It does not mean those products add nothing, and we do not say what they do. Canva's AI terms name C2PA metadata in a rule for users, but the two Canva pages we read do not say that Canva adds it; our guide to Canva's images has the details. The table also leaves out visible marks.
What do these signals survive?
It depends on the signal and on what happens to the file. OpenAI says Content Credentials can be lost through everyday actions, such as taking a screenshot or converting a file. It says SynthID is designed to withstand common changes, but that heavy cropping or file compression can make a watermark undetectable. Midjourney says its tag is easily lost, and names screenshots, edits and most social media sites. So a file you meet online may have lost what the table describes. The Content Credentials explainer covers the signed record, and What is SynthID? covers the watermark.
What did our checker read in files we hold?
We keep 32 real files: 29 downloaded from Wikimedia Commons categories named "Images generated by" a product, 2 vendor samples (Adobe Firefly and ChatGPT) and 1 C2PA test image. Nine of the 32 carry a signed Content Credentials record. Those nine are the 3 files labelled Adobe Firefly (2 filed under Firefly, 1 vendor sample), the 2 files filed under Image Creator by Microsoft Bing, 3 of the 4 ChatGPT files (2 filed under ChatGPT, 1 vendor sample) and the C2PA test image. The other 23 carry no record. One of the nine, a ChatGPT file, reads as trusted on the current C2PA Trust List. The full tables are on our accuracy page.
Read this with care. Third parties uploaded the Commons files, and we do not know how each file was handled before upload. The files show what the checker reads in files we hold. They do not show what a product does. For example, the 2 files filed under Flux carry no record, and we cannot say why. Black Forest Labs says it attaches credentials to output from its own API and hosted tools. The checker also does not read SynthID or Midjourney's hidden tag, so a file can carry either and still show nothing from us.
What this means when you check an image
Open the checker, choose the file and read the headline. Three headlines come up most often for files from the makers above.
- "Content Credentials found; legacy trust-list match". A signed record is present and its signer is on the frozen Interim Trust List, not the current C2PA Trust List. The Firefly and Microsoft files read this way. It tells you who signed the record. It does not tell you the record is true.
- "Content Credentials found; signer trust not established". A signed record is present, but neither list we load vouches for the name in its certificate. Two of the ChatGPT files read this way. A record made with a self-made certificate reads the same, so this headline cannot separate the two.
- "No supported provenance signals found". The checker found none of the signals it reads. This does not mean the image is real or human-made, and the file may still carry a watermark we do not check.
Every result also says that these checks cannot determine whether an image is AI-generated. For SynthID, use the maker's own checker, as What is SynthID? explains. For a broader view of all the signals, see How to tell if an image is AI.
Guides on this topic
- Does ChatGPT add a watermark or metadata?
- Gemini and Nano Banana: SynthID and credentials
- Does Midjourney add a watermark or metadata?
- Microsoft Image Creator and Content Credentials
Frequently asked questions
- Do all AI images carry a watermark?
- No. The makers' own pages describe different signals for different products, and several say coverage varies. Files also lose signals when they are screenshotted, converted or shared.
- If a file has no Content Credentials, does that mean it was not made by an AI tool?
- No. Finding nothing does not show how an image was made. The signals may never have been added, or they may have been lost along the way.
- Does the New Watermark checker read SynthID?
- No. We do not check SynthID, Meta's watermark, Adobe TrustMark or Midjourney's hidden tag. Every result has a "Not checked here" line that names SynthID, Meta's watermark and Adobe TrustMark, with links to some of the makers' own checkers.
- Why do some Content Credentials files say the signer is not trusted?
- Trust means a list we load vouches for the signer's certificate. Two of the ChatGPT files we hold carry expired certificates that no list we load trusts, so they read as "signer trust not established". A third is trusted on the C2PA Trust List.