ChatGPT Watermark Remover: What Works

There is nothing to remove. OpenAI has said it developed a text watermarking method for ChatGPT and has never deployed it, so ChatGPT text carries no watermark today and no tool sold as a "ChatGPT watermark remover" is removing one. The model that genuinely does watermark its text is Claude, and Anthropic confirmed again on August 29, 2026 that no public detector for it exists yet. What is actually flagging your writing is a statistical detector, which is a different problem with a different fix.

This search term is one of the most common in the category and it rests on a premise that is not true. That matters, because people are buying products to solve it. Here is what each of the major model makers has actually shipped, what a "remover" can and cannot touch, and what to do about the thing that is really getting your work flagged.

Does ChatGPT watermark its text?

No. OpenAI has publicly described a text watermarking method it built and has not released, and it has published its reasoning for holding it back, including the concern that watermarking "could stigmatize use of AI as a useful writing tool for non-native English speakers." A method that exists in a lab and has never been switched on leaves no trace in your document. If you paste a ChatGPT answer into a file today, there is no hidden statistical signature in the word choices for anyone to find.

The confusion is understandable, because OpenAI does mark some things. Images generated by its models carry C2PA provenance metadata, which is a signed record attached to the file rather than anything embedded in text. That is a real, shipped mechanism, and it is completely unrelated to the words in an essay. We wrote up the full first-party account of what OpenAI has and has not deployed in the ChatGPT watermark.

Which AI models actually watermark their text right now?

Compiled from each maker's own published material. Anthropic's help centre re-read August 29, 2026.

The Claude row is the one that changed the picture this year. Anthropic began marking all text from supported Claude models at the model level, which means it applies everywhere Claude is used, travels with the text when it is copied and pasted, and "may persist through some editing." That is a genuine, shipped watermark on real output people are submitting every day.

And yet it still cannot be checked. Anthropic's own help centre says it is "working to enable users and other third parties to detect Claude's embedded watermarks" and that it will "share details on detection mechanisms in forthcoming technical documentation." It also cautions in the other direction that a detected mark "provides a signal that content was processed by Claude, but is not fully conclusive." We keep a running account of that on Claude AI detector.

Do AI watermark removers work?

The honest answer is that nobody can tell, and that is the reason to be sceptical rather than reassured. A claim that a tool removes a text watermark is currently untestable in both directions: with no public reader for any major model's text mark, you cannot show a watermark is present before the rewrite, so you cannot show it is absent after. A vendor making the claim is not necessarily lying. They are describing a demonstration that does not exist.

Phrasly is the clearest example, because it puts the claim at the top of its pricing page: "The only humanizer that completely removes all AI watermarks." Its own FAQ then makes a much better and more defensible argument, which is that it runs its own models rather than wrapping OpenAI's, so its output never passes through a system that could watermark it in the first place. That is future-proofing rather than removal, and it is a real difference. We went through the whole thing, including what its plan actually costs, in our Phrasly AI review.

What is actually flagging your text, if not a watermark

A detector. Turnitin, GPTZero, Copyleaks, Originality.ai and ZeroGPT do not look for a hidden mark, because there is none to look for. They measure statistical properties of the writing itself: how predictable each word is given the ones before it, and how much that predictability varies from sentence to sentence. Human writing wanders. It runs a long sentence into a short one, picks an odd word, changes register mid paragraph. Model output is smoother, and smoothness is what gets scored.

This is why the two problems need opposite solutions. A watermark, if one existed, would be a signal to strip. A detector score is a property of your prose, and the only thing that changes it is changing the prose. There is no metadata to delete, no file header to rewrite, no setting to toggle.

It is also why detector scores are unreliable enough to be worth understanding before you panic about one. These classifiers are trained on enormous corpora of human and machine text, most of it gathered by scraping the open web at scale, and the labels in that training data are far messier than the confident percentage in the output suggests. Non-native English writers get flagged at dramatically higher rates than native speakers, which is exactly the harm OpenAI cited when it declined to ship its watermark. Our breakdown of how often detectors are simply wrong is on the AI detector false positive rate.

What about the images? That part is real

If you generated a picture rather than text, provenance metadata is a genuine, shipped thing. OpenAI attaches C2PA credentials to images from its models, and Anthropic attaches signed provenance metadata to files where supported. Unlike a text watermark, this is a normal piece of file metadata, and it is fragile in the ordinary course of business: it does not survive a screenshot, and many platforms strip metadata on upload as a matter of routine, without anyone intending to remove anything.

So an image-focused "AI watermark remover" is at least operating on something that exists. It is just a completely different mechanism from the one people mean when they worry about an essay, and the two have been collapsed into one search term by tools that sell against both.

What to buy instead, if anything

Start by deciding which problem you actually have, because they need different things.

If a detector flagged writing you wrote yourself, do not buy a rewriter. Rewriting your own honest work to please a classifier makes it worse and destroys the one thing that will actually clear you: a version history showing the document being built. Drafts, comments, timestamps and a paper trail beat any tool here. We set out what to do, in order, in how to prove you did not use AI.

If you are drafting with AI and want prose that reads like a person wrote it, that is what a humanizer is for, and it is worth being precise about what you are buying. You are buying a rewrite that restores variation in sentence length, rhythm and word choice. You are not buying a guarantee, because no vendor controls what Turnitin does next month, and any vendor promising one is promising something outside their control. The paid field, priced and compared on what each one will actually commit to, is on best AI humanizer.

If you are choosing which model to draft with in the first place, the watermark situation is now a real input into that decision, and it is the one genuinely new consideration this year. Claude output carries a mark. ChatGPT output does not. Neither is currently readable by anyone you are likely to be submitting to, but only one of them has a mark that could become readable later, on text you have already handed in. We compare the models on this in which AI is least detectable.

The short version

ChatGPT does not watermark text, so a ChatGPT watermark remover is a product for a problem that does not exist yet. Claude does watermark text, and no public tool can read that mark, which means no tool can prove it removed one either. Image provenance metadata is real, separate, and fragile in normal use. And the thing that flagged your document was almost certainly a statistical detector reading your prose, which no remover touches. Fix the prose or prove the process. Those are the two options, and neither of them is sold as a watermark remover.