A ChatGPT watermark is OpenAI's Content Credentials: a visible CR symbol plus invisible C2PA metadata embedded in the image file. This tool does not remove it, and it will not pretend to. What this tool does is narrower, and we put it first rather than at the bottom: it removes the visible Gemini sparkle from a Gemini image, and only that. If your file came from Gemini, that is the page you want; if it came from ChatGPT, the rest of this page tells you what is actually on it and why a pixel tool is not the answer.
A visible symbol and a hidden record
OpenAI announced in February 2024 that images made with DALL·E 3 — through ChatGPT and through the API — would carry C2PA Content Credentials. That credential is not one thing; it is two, kept in different places.
- A visible CR symbol. The Content Credentials mark, drawn onto the picture so a person can see it. It is made of ordinary picture content, and it sits in a fixed place, much like any badge.
- Invisible C2PA metadata. A signed record stored inside the file's container, alongside the pixels rather than within them. It is the part a detector reads and a person never sees.
Both are real. Both are on the file. They simply live in different layers, and that difference decides what any tool can and cannot reach.
The single job this tool is built for
The tool here holds exactly one template: the visible badge of a single generator — Gemini — in the two layouts that generator has used. Its work is a reversal of the placing operation inside that one rectangle, run in your browser, with your image never uploaded. That is the whole of it.
So, for a ChatGPT image, two sentences cover it. It does not remove the CR symbol, because that mark is not the badge the tool holds a record of. And it does not touch the C2PA metadata, because that record is not made of pixels at all. Neither is a limitation we are working around; both are consequences of what the tool is.
Why a different generator's symbol is out of reach
Reversing a badge is possible only when the mark is a known template rather than an unknown one. A template is a measurement of one object at one size in one position, and a tool can only match the record it was given. A CR symbol from OpenAI is a different object, so it is a different record — and no record of it is loaded here. The first step of the method is matching the stored template against the mark in the picture; with no template, there is nothing to match. The input is missing, not merely hard to work with.
Why the metadata is beyond a pixel tool entirely
The C2PA manifest is not picture data. It is a record in the file's container, and a tool that rewrites pixels does not read or write that container. This is why cleaning a visible mark — of any kind — would leave the metadata exactly where it was: the two were never in the same layer to begin with. A file can therefore show a clean surface and still carry the record, and that is not a contradiction; it is two separate systems doing two separate jobs.
If your file is a Gemini image
Then the tool is pointed at the right thing. It takes a still in PNG, JPEG or WebP, or a clip in MP4, WebM or MOV, and it works on the visible Gemini sparkle alone — nothing else in the frame, and no metadata, is altered. The Gemini watermark page sets out the cases it does and does not cover, and the tool on the front page is where a file of your own can be tried against it.
Two neighbouring write-ups take the same distinction from other angles: the wider term is untangled on a page that keeps its two senses apart, and a different company's mark, kept in four places at once, is worked through at Meta AI.
Sources
- OpenAI, C2PA in DALL·E 3 (help.openai.com), and OpenAI's announcement of 7 February 2024. Source of the statement that images generated with DALL·E 3 — in ChatGPT and via the API — carry C2PA Content Credentials, comprising invisible metadata and a visible Content Credentials (CR) symbol. Recorded with its date: it describes the position at announcement, not a measurement of any file today.
- Coalition for Content Provenance and Authenticity (C2PA), the open technical standard for signed content provenance. Source of what a C2PA manifest is and where it is stored — in the file's container, not in its pixels.
- Our own implementation, read on this site. Source of the scope stated above: one badge, two layouts, a still or a clip, no metadata handling, and no capture of any other generator's mark.
- Our own observation of the results returned for this term, checked 25 October 2026. Recorded as a pattern only: no site is named, no ranking position is claimed, and no page's wording is reproduced.