Removal, precisely

Can you remove a ChatGPT text watermark?

You can remove the things people confuse with a ChatGPT text watermark. You cannot delete textGrain the way you delete a character or a metadata box, because textGrain is the choice of words. Changing enough words can make the pattern harder to detect. That is a rewrite, not a scrub, and it does not certify anything about authorship.

This page is for people who already have a draft they are allowed to edit and need the layers separated. It is not a method for slipping past a school, a client, or a disclosure rule. OpenAI’s description of the mark is in the provenance help. The mechanism is in how textGrain works.

Three operations called removal

Products that advertise a single “remove ChatGPT watermark” button are collapsing this table. If the button rewrites the paragraph, it is the third row. If it only deletes U+200B, it is the first. If it re-saves a JPEG, it might be the second, and only for the manifest. Ask which row before you trust the past tense “removed.”

Characters you can delete

Zero-width spaces, joiners, a BOM, and bidi controls are real and removable. The ChatGPT watermark detector counts them and can delete the listed ones without rewriting a sentence. That is worth doing when a command fails, a diff is noisy, or a CMS stores two “identical” titles. It is described in ChatGPT hidden characters.

OpenAI’s help text says the text watermark does not add those characters, and that copying does not introduce hidden material as the mark. So a successful clean is evidence about the clipboard, not evidence that textGrain is gone. Re-checking until the count is zero is the whole of the proof available on this layer.

Manifests you can strip

A PNG or JPEG can carry C2PA Content Credentials and AI-looking XMP even when you only meant to publish a picture. Stripping that container mark is a file operation. This site’s images tab does that for the segments it recognizes, and leaves pixels alone. SynthID in the pixels is a different signal and is not removed by dropping APP11. The boundary is the point of the C2PA guide.

Document properties on PDF and DOCX are closer to the manifest row than to textGrain. You can clear a generator tag and still have every sentence marked in the statistical sense, or the reverse: unmarked sentences inside a file that still says algorithmic media because nobody touched the properties. Check the container after you edit it. Do not infer the sentences from the tag, or the tag from the sentences.

Stripping a credential to hide origin where a disclosure rule requires the mark is the use this site tells you not to make. The same bytes are legitimate to remove when they are leftover on a file you authored, or when you are archiving your own export and the manifest is the wrong record. Know which case you are in before you click Clean.

Wording you would have to change

textGrain is tested across many token positions. OpenAI says light edits are in the set of changes the watermark is designed to withstand, and that substantial paraphrasing or translation can make detection less reliable. There is no published public recipe of “change these twelve function words and the key will miss,” and this page will not invent one. The direction is enough: the pattern is the wording, so only a real change to the wording can touch it, and even then the result is a weaker signal, not a receipt.

A humanizer or spinner is a second model doing that rewrite. It may also be exactly what a classifier has been updated to notice. Turnitin’s AI writing report includes likely AI-paraphrased text. You can blunt one measurement and brighten another in the same pass. Neither outcome is yours to confirm without those vendors’ tools. See watermark versus an AI detector.

Short text and code were never strong carriers. If the passage is under the range OpenAI discusses, around 200 tokens for the code-of-practice floor, “removal” is the wrong word. There may have been too little signal to detect in the first place. That is not a cleaned watermark. It is a specimen the scheme does not claim.

What a quiet detector still does not prove

Suppose you later obtain text-detector access and the score is under the threshold. OpenAI’s own list of reasons includes text from before the rollout, an unsupported path, an API project that never opted in, a short or factual answer, code, and text that was rewritten. A quiet result is compatible with all of those, including “a model wrote this and the mark did not survive.” It is not a finding that a person wrote it. The longer version is what a missing watermark means.

This site cannot produce that quiet result for textGrain, because it does not have the key. If a page promises to show you the watermark disappearing, it is showing you a different instrument, usually a classifier moving, or a character count moving.

When removal is the wrong goal

Use the character cleaner and the metadata cleaner on drafts and files you own when the problem is a broken paste or a stale tag. Keep the credential when the context requires a machine-readable mark. Do not treat a rewrite as a way to tell a reader, a client, or an instructor that a model was not involved. Disclosure is a sentence you write. It is not a side effect of an eraser.

If the practical question is whether your writer’s file will announce itself, start with the check, not the clean button. The sequence for editors is in how to check a blog writer. Which free tools even belong in that sequence is in the checker comparison.

Three files, three results

The same afternoon can produce three “please remove the watermark” requests that should not share a button.

A tutorial paste. A reader copied a command from a chat and the shell says command not found. Check finds a zero-width space or a BOM. Clean deletes that codepoint. The command is byte-for-byte the visible command. Re-check returns zero. textGrain was never the bug. If the command was also generated in a marked EU session, the wording pattern is still in the line, and a one-line command was a weak carrier anyway. You fixed the terminal. You did not file a provenance report.

A header image. A PNG from ChatGPT images still has a C2PA manifest. The images tab reports it. Clean drops the container segments it recognizes. The pixels are the same picture, which means a SynthID signal in those pixels is the same signal. openai.com/verify remains the OpenAI surface for that pixel and audio question. Content Credentials Verify will now miss the manifest you removed. Write down that you removed it, and do not describe the PNG as unmarked in every sense.

A 900-word post. Someone asks for the text watermark to be gone and for the sentences to stay. Those constraints contradict each other. textGrain is the sentences. A light copy-edit leaves the pattern in the set of edits OpenAI says the scheme is built to survive. A full rewrite may drop detection and will also change the piece you thought you were publishing. There is nothing to re-check on this site afterward except the character layer, which was never the question. If a style score moves, that is the classifier reacting to the new wording, documented in watermark versus an AI detector.

What humanizer pages are selling

Search results for “remove ChatGPT watermark” are dominated by paraphrasers. The pitch is a percentage that falls. The percentage belongs to a classifier the page does not name precisely, measured on a sample you cannot audit. Nothing in that flow shows textGrain before and after, because the page does not have the key. Paying for the rewrite does not buy the measurement.

The rewrite can also be self-defeating. Turnitin’s model now treats likely use of an AI paraphraser as part of the AI writing report. A client-side “humanizer” that keeps the facts and swaps the connectives is the pattern that update describes. You may leave a campus report louder than you found it. That is not a reason to learn a bypass. It is a reason to stop treating the spinner as a watermark eraser. Institutional scoring is covered in can Turnitin detect ChatGPT watermarks.

A second sales claim is “we strip hidden Unicode, so the watermark is gone.” The first half can be true. The second half is the category error this page exists to prevent. If a vendor will not say which row of the table they implement, assume they are selling the row that is easiest to demo.

When the mark should stay

Removal is the wrong default for teams who turned watermarking on deliberately. An API project that opted in did it so downstream software could tell the text had passed through the model. Rewriting the output in a second undocumented pass, or routing around the setting, fights the reason the switch exists. If you are the provider’s customer, the setting lives under organization data controls and under the project override. Record when it was enabled and which models were selected. Detector access is still separate. You can mark text you cannot later score.

Publishers sit one step downstream. If a disclosure rule or a platform wants the machine-readable mark, stripping C2PA so the badge disappears is a compliance choice, not a cleanup. The EU side of that choice is what the AI Act asks for. Hygiene remains available for the accidents: a BOM in a code block, a generator key left in Markdown you wrote by hand, a manifest on a scan of a page you authored. Run Check, read the carrier name, and clean only that carrier.

Keep the original next to the cleaned file for a while. The only honest sentence later is “we removed these listed characters” or “we removed this manifest.” The sentence “we removed the ChatGPT watermark” is the one the table does not support unless you changed the wording and you are willing to say that the wording changed. If a teammate needs the coverage map before they decide the mark was ever present, send them does ChatGPT watermark text in 2026 rather than a before-and-after screenshot of a style score.

If you are documenting the work for someone else, attach the before and after files and name the carrier in the filename note. “cleaned-zwsp” and “stripped-c2pa” are descriptions a later reader can test. “dewatermarked” is not. The index of the other guides, if they need the neighboring question, is the blog hub.

FAQ

← All AI watermark guides