Since 20 May 2026, an image made in ChatGPT carries two marks where it used to carry one. The first is C2PA metadata, which OpenAI has added to its images since February 2024. The second is a SynthID watermark in the pixels, the same technology Google uses, which OpenAI announced that day for images from ChatGPT, Codex and the API. So yes, ChatGPT adds metadata to images. It also adds something a metadata remover can’t reach.
The metadata part, from DALL·E 3 on
OpenAI started signing DALL·E 3 images with C2PA in February 2024, in ChatGPT and through the API, with mobile following on 12 February. Images made inside ChatGPT got a second manifest on top, recording that ChatGPT made the request, which PetaPixel described at the time as a “dual-provenance lineage”. Nothing about it is visible. The small “CR” badge in PetaPixel’s report appears only when the file is opened on the Content Credentials Verify site.
OpenAI was frank about how fragile this layer is. Its announcement said C2PA metadata “can easily be removed either accidentally or intentionally”, that most social platforms strip it on upload, and that a screenshot removes it too. That is still true, and it is why the May 2026 change matters more than the 2024 one.
What changed in May 2026
OpenAI added Google’s SynthID watermark to its images and opened a public checker at openai.com/research/verify. According to OpenAI’s description, quoted by PetaPixel, the tool looks for “a SynthID watermark that originates from OpenAI, or a trusted C2PA manifest that originates from OpenAI.” Either one is enough. OpenAI’s reasoning for using both is that “watermarking can be more durable through transformations such as screenshots, while metadata can provide more information than a watermark alone.”
OpenAI also says of the checker: “Errors can happen, but they are rare.” We haven’t seen a false-positive or false-negative figure from OpenAI to go with that sentence, and until there is one, “rare” is a promise rather than a measurement.
What cleaning a ChatGPT image changes, and what it leaves
A metadata clean removes the C2PA manifests and anything else in the file’s metadata blocks. In a JPEG that is the APP11 segments; in a PNG it is the caBX chunk; in a WebP it is the C2PA chunk. WipeTheAI does this losslessly on your device as its first step and lists what it found, but both of its wipe modes then change the pixels, so there is no metadata-only download. If metadata is all you want gone from a JPEG, exiftool -jumbf:all= image.jpg deletes the manifest and leaves the image data alone. For an image made before May 2026, the manifest is most of what the file had to say.
For anything newer, it isn’t. The SynthID mark sits in the pixel values, so a metadata-only clean leaves it exactly where it was. Going by OpenAI’s description, the verify page should still recognise a metadata-stripped image through the watermark alone. We haven’t run a stripped-only file through it, so treat that as a reading of OpenAI’s words, not a result.
Disrupting the watermark means changing the pixels. The mode built for that, Maximum disruption, regenerates the image lightly two times over. On our test image the result passed OpenAI’s verify check. That is one image. We tuned the recipe against Gemini’s provenance check (a classifier judgement, not Google’s production detector), and per-image results vary; the page on what we measured with SynthID has the runs, including the ones that failed. Maximum disruption uploads your image to our servers, and the tool warns you before it does.
Who this is for
Mostly people who made something in ChatGPT, edited it further, and want to know what the file says before it goes into a client deck or a portfolio. If your aim is to present a ChatGPT image as a photograph, we are not the tool for that, and our acceptable use policy says so. For the question of whether removing it is allowed at all, see is removing AI metadata legal.