ChatGPT images have no visible watermark to remove. Since 19 May 2026 they carry two invisible layers instead: C2PA metadata in the file header, and DeepMind's SynthID embedded in the pixels. Tools sold as ChatGPT watermark removers advertise removal of the first one.
We do not offer a watermark remover — there is no watermark-removal model on our provider shelf. What follows is what the two layers are, and where each one lives.
Last updated: July 2026
TL;DR
- There is no visible watermark on ChatGPT images. No burned-in badge at any tier. The mark people picture when searching for a remover does not exist.
- Two invisible layers do exist: C2PA metadata in the file header, and Google DeepMind's SynthID embedded in the pixels. Both apply to free, paid and API output.
- A remover strips the metadata layer. C2PA lives in the header and comes off — re-saving a file can drop it by accident.
- SynthID is in the image content, and DeepMind describes it as designed to survive "cropping, adding filters, changing frame rates, or lossy compression".
- This is the mirror image of Sora, where the visible logo was on free exports and C2PA turned up only on the clean paid ones.
What a ChatGPT watermark remover actually removes
Three watermark layers, and only the middle one is in play.
A visible watermark — not present. ChatGPT does not burn a watermark badge into image output. If you arrived here expecting to erase a corner mark, there is nothing there. This alone resolves most searches for a ChatGPT watermark remover.
C2PA metadata — the layer removers advertise. A provenance manifest carried in the file header rather than in the picture. Tools in this category describe themselves precisely — "remove C2PA metadata in your browser" — and that is a header edit, which is the kind of operation a file rewrite can perform. We have not tested any of these tools, so we are describing what they advertise rather than what we measured.
SynthID — stays. OpenAI integrated DeepMind's SynthID watermark alongside its C2PA adoption. SynthID is not metadata. It is embedded in the content itself, and DeepMind's own description is that the watermark is "added the moment content is created, and designed to stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression."
So a ChatGPT watermark remover that advertises "removes C2PA metadata" is describing its work accurately, and that work does not touch the pixel-level layer.
Why a ChatGPT watermark remover differs from one built for Sora
The two OpenAI products watermark in opposite directions, which is why advice written for one misleads for the other.
With Sora, the visible logo appeared on free and Plus exports, and independent testing found C2PA metadata only on the watermark-free Pro downloads — so the file that looked branded was the unsigned one. We covered that in our piece on Sora watermark-free downloads.
ChatGPT images invert it: nothing visible on any tier, provenance on every tier, and one of the two provenance layers designed specifically to survive editing. Sora's watermark was tier-dependent. The ChatGPT watermark is not.
The practical consequence: with Sora, paying removed the visible watermark. Here, paying changes nothing about the watermark at all.
Does a watermark remover make an image undetectable?
We cannot answer that from what we verified, and the honest version is more useful than a reassuring one.
Stripping the header removes one watermark signal. The pixel-level watermark is a different mechanism in a different place, and DeepMind's stated design is that it withstands cropping, added filters, frame-rate changes and lossy compression. We have run no detection tests, so we will not tell you where either layer fails. The one structural point we can make: they are independent, so an operation on the header says nothing about the pixels.
There is also a category of detection that has nothing to do with watermarks at all: statistical and forensic analysis of the image itself, which predates both standards and reads the picture rather than its header. We did not evaluate how well it works.
What we will say is that "I removed the metadata" and "this cannot be identified as AI-generated" are different statements, and only the first one is about something you did.
Is using a ChatGPT watermark remover legal?
The framing that matters is not the watermark removal itself, it is what you do afterwards.
DMCA §1202 in the US governs the removal or alteration of copyright management information. Stripping provenance data to facilitate infringement or to conceal origin deceptively is where exposure begins, and it carries statutory damages rather than a takedown notice.
Platform disclosure rules are a separate obligation, and no watermark remover touches them. Where a platform requires AI-generated content to be labelled, that requirement is about what you declare, not about what the file contains — so editing the file does not answer it either way. Check the policy of wherever you are publishing; they differ.
The manifest can work in your favour too. C2PA is a provenance standard — its purpose is recording where a file came from, which cuts both ways. If the origin of an image ever comes into question, an untouched original is the copy that still carries that record. Keeping one costs nothing, so edit copies.
What we do instead of a watermark remover
Nothing, on the removal side — there is no watermark-removal model on the shelf we buy from. What we do is generate. Across the four Nano Banana 2 runs we did for a separate prompt test, the output files carried no visible watermark; rates are 8 credits per 1K image, 12 at 2K, 18 at 4K, shown before every run. Starter is $248/year for 33,500 credits.
One caveat we will not skip: "no visible watermark" is not "no watermark". We did not inspect those files for invisible provenance data, and the sensible default with any AI-generated image is to assume some may be present.
FAQ
Does ChatGPT put a watermark on images?
Not a visible watermark. Since 19 May 2026 it applies invisible provenance: C2PA metadata in the file header plus DeepMind's SynthID embedded in the pixels. Both are applied to free, paid and API output alike.
Do free and paid ChatGPT images have different watermarks?
No — the marking is the same across tiers. This is different from Sora, where the visible logo appeared on free and Plus exports but not on Pro downloads.
What does a ChatGPT watermark remover actually do?
It strips C2PA metadata from the file header. That is a real operation on a real watermark layer. It does not address the SynthID watermark, which DeepMind describes as embedded in the content and designed to survive cropping, filters and lossy compression.
Can the SynthID watermark be removed?
We have not tested it and will not claim a number. DeepMind states it is designed to withstand cropping, added filters, frame-rate changes and lossy compression. Treat it as durable rather than as something a metadata tool addresses.
Will a watermark remover stop AI detectors?
A header edit and a pixel watermark are independent, so removing the first does not act on the second. Separately, forensic analysis reads the picture rather than the header. We tested no detector, so we are describing which layer each thing touches, not predicting an outcome.
Should I keep the original file?
Yes, if the image matters commercially. C2PA is a record of where a file came from, and only the untouched copy still has it. Edit a duplicate.
Sources
- OpenAI's adoption of C2PA and integration of Google DeepMind's SynthID across ChatGPT, the OpenAI API and Codex, dated 19 May 2026 — every generated image is marked
- Google DeepMind SynthID technical description — watermark "added the moment content is created, and designed to stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression"; embedded in content rather than metadata
- gpt-image-2 tier behaviour, July 2026 — file-level provenance metadata plus content-level pixel watermark on every tier, free and paid alike
- Our own reporting on Sora's tier-dependent marking, including independent C2PA testing showing the manifest on watermark-free Pro downloads only
- DMCA §1202 — removal or alteration of copyright management information; statutory damages
- src/shared/lib/credits.ts and kavel.ai/pricing — 8 / 12 / 18 credits at 1K / 2K / 4K; Starter $248/year = 33,500 credits; cost preview before each run
- ~/.claude/pricing shelf check — no watermark-removal model available to this site
