Answer
What metadata do ChatGPT images contain?
Principally C2PA content credentials — a signed record marking the image as AI-generated — alongside ordinary format-level fields such as dimensions and colour profile.
Updated
The significant part is the content credential. It is a cryptographically signed manifest embedded in the file, asserting that the image was produced by a generative model and identifying the issuer. Unlike EXIF, it is signed, so altering it is detectable rather than silent — which is the entire point of the scheme.
Where it sits depends on the container. In a JPEG the credential lives in an APP11 marker segment; in a PNG it is a dedicated chunk. Neither location is somewhere ordinary metadata viewers look, which is why an image can report almost no EXIF while carrying a full provenance manifest.
Alongside that you get the routine fields any encoder writes: dimensions, colour profile, sometimes a software identifier. There is no camera data, because there was no camera, and no GPS, because there was no location fix.
Locally-run generators are a different matter and frequently leak far more. Stable Diffusion front-ends write the complete prompt, negative prompt, seed, sampler, CFG scale, model name and hash into a PNG text chunk in plain readable text. If a prompt represents work you would rather not hand over, that is the file to check before sharing.
Removing any of this is straightforward, and it is worth being clear about what it achieves: it does not make the image undetectable as AI. It removes a label, not the underlying signal. The ChatGPT image metadata remover handles the removal, and the C2PA explainer covers what credentials are for.
More on this: ChatGPT image metadata remover, AI image metadata remover.
Terms used here: EXIF, C2PA, PNG text chunk, Chunk.