Every supported image ChatGPT makes carries two provenance signals you can't see: C2PA metadata, which records that OpenAI's tools made the file, and a SynthID watermark embedded in the image itself. So ChatGPT does add watermarks to images, and a visible one appears only when you ask for it. Below is what each signal records, how to check a file, how to add a visible label on purpose, and what these signals can't tell you.
Does ChatGPT add watermarks to images?
Yes: two invisible ones, and a visible one only on request. OpenAI's help center says supported images from ChatGPT, Codex and its API carry C2PA metadata as well as a SynthID watermark, and the ChatGPT Images 2.5 system card, published September 8, 2026, lists both among its provenance tools. As of September 2026:
| Signal | What it is | Visible on the image | How to check it |
|---|---|---|---|
| Visible watermark | A logo or line of text drawn into the picture | Yes, but only when your prompt asks for one | Look at the picture |
| C2PA Content Credentials | Signed metadata recording which tool made the file | No | OpenAI's Verify page, or the Gemini app on the web or Android |
| SynthID | An invisible watermark embedded in the image content | No | OpenAI's Verify page |
OpenAI notes that which signals a file carries can differ by product and model, by how and in what format the file was exported, and by when it was made. An image made in ChatGPT this month and one made two years ago may not carry the same signals.
What C2PA Content Credentials record
The C2PA standard, from the Coalition for Content Provenance and Authenticity, is an open specification for attaching provenance information to a file, and Content Credentials is the name for that information. OpenAI says the metadata in its images can include the tool that made the file, when it was made, and other details of its history. The standard isn't only for AI: camera makers and news organizations are adopting it to show where real photos and footage came from.
- It's tamper-evident. The C2PA says each credential is digitally signed and includes cryptographic hashes of the file, so any change to the file or its metadata breaks the link.
- It can go missing. The C2PA and OpenAI both note that metadata can be removed, on purpose or by accident, when a platform, editing tool or file conversion drops it. A missing credential proves nothing either way.
- It doesn't name you. The C2PA's core specification doesn't support attributing content to individuals or organizations, and OpenAI says a verification result doesn't identify who created an image or why.
- It's widely backed. Content Credentials has grown into a collaboration of more than 500 companies, with Microsoft, Adobe, Sony, the BBC, OpenAI, Google and Meta among those leading it.
What the invisible SynthID watermark adds
ChatGPT does add invisible watermarks, and the technology behind them is SynthID, Google DeepMind's watermarking system for AI-generated content. SynthID embeds a signal people can't see into the image itself, and Google says it's designed to stand up to changes like cropping, filters and lossy compression. The ChatGPT Images 2.5 system card describes adding SynthID watermarking across ChatGPT, Codex and the OpenAI API to make its provenance more resilient.
The two signals back each other up. OpenAI's help center explains that if a file's metadata is gone, the embedded watermark may still show that the image came from its tools. Metadata carries more detail, such as which tool made the file, while the watermark carries less information but stays part of the picture. The C2PA calls this pairing of a credential with an invisible watermark or fingerprint "Durable Content Credentials".
The checker matters, though. Google's Gemini app can read SynthID, but Google says it currently recognizes only content made by Google's own AI tools, even though other companies have started using SynthID. To check for OpenAI's watermark, use OpenAI's own tool.
How to check if an image was made with ChatGPT
OpenAI's Verify page checks a file for both of its signals.
- Open OpenAI's Verify page.
- Upload a single image in PNG, JPG or WEBP format. It also accepts audio files.
- Use the file as you received it. OpenAI says not to crop the image or save it in a different format, such as PDF, before you check.
- Read the result. It reports whether it found a SynthID watermark, C2PA metadata, both or neither.
A detected signal means the image likely came from OpenAI's tools, and OpenAI says false positives are rare. It doesn't tell you whether the picture is accurate, who made it, or how it has been used since. No signal doesn't mean a person took the photo: the metadata may have been stripped, the watermark degraded, the image may predate these signals or come from a legacy model, or another company's AI may have made it. OpenAI says uploaded files aren't stored unless legally required and aren't used to train its models.
For Content Credentials from any maker, the Gemini app summarizes what a file's credentials say, including edit history and whether AI was used, on the web and on Android. For clues you can see with your own eyes, such as garbled signs and extra fingers, use this checklist for spotting AI images.
How to add a visible watermark in ChatGPT
ChatGPT does create a watermark you can see, but only when your prompt asks for one. OpenAI's help center gives the example "Include a visible OpenAI watermark in the image." You can also write your own label, such as "Add a small white 'AI-generated' label in the bottom-right corner." A visible mark is separate from the hidden signals: it helps people who look at the picture, while software reads the metadata and the watermark. Check the spelling in the finished image, because AI-drawn letters can come out wrong. The steps for making and editing pictures are in the guide to making images with ChatGPT.
Does ChatGPT add watermarks to text or audio?
Audio, yes. Text, not yet. OpenAI's help center lists SynthID for supported audio, where the watermark is inaudible and built into the sound itself. For text, OpenAI says its goal is to extend provenance signals to all content types, including text, as part of its commitments to the European Commission's Code of Practice on Transparency of AI-generated content. As of September 2026, its table of covered content types names images and audio only. Google, by contrast, says it already applies SynthID to text generated in the Gemini app.
What the hidden watermark means for your ChatGPT images
The signals travel with the file, so they affect where and how you post it.
- Social media labels. Meta labels content "AI info" on Facebook, Instagram and Threads when it detects industry-shared signals or the poster discloses AI use. Since a September 2024 update, content that AI only edited gets the label in the post's menu instead.
- Marketplaces. Rules like Etsy's AI disclosure apply whether or not a file carries a watermark. Etsy's rule and others are covered in the guide to selling AI art.
- Ownership. A watermark records origin, not rights. The guide to who owns an AI image explains what copyright can and can't protect in a ChatGPT picture.
- Trust. OpenAI says a provenance signal can't vouch for whether a picture is true or unedited, who legally owns it, or whether it's shown in the right context.
Watermarks on images you make in Toybox
Toybox AI doesn't currently add watermarks to AI-generated results. Memes made on the free plan include a small Toybox logo. Toybox doesn't add content credentials (C2PA) or any provenance label of its own. Some AI providers build invisible watermarks or credentials into what their models make, and Toybox doesn't control or document those.
If you want viewers to see that a picture is AI-made, put the label in the prompt. In Image Generator, write the label text in double quotes and switch on "Enhanced quality" under "More options", which the tool describes as sharper text and logos (it adds 45 credits to the current 50):
A watercolor illustration of a red lighthouse on a rocky coast at sunrise, pink and gold sky, gulls over the water. In the bottom-right corner, small white sans-serif text that reads "AI-generated".Zoom in on the label before you post the picture, and create the image again if a letter is wrong.