How to Check If ChatGPT Added a Watermark to Your Text

There is no consumer tool that reads ChatGPT output and reports a watermark. OpenAI has not shipped a text watermark, and no detector exposes a watermark check. What you can do is inspect the text for invisible characters, read OpenAI’s own documentation, and fix the patterns that get your draft flagged. That is the honest answer, and the rest of this guide is the workable part.

Why people believe ChatGPT watermarks text

The belief has a real origin. In the spring of 2025, users spotted unusual invisible Unicode spacing characters in ChatGPT output and declared the secret ChatGPT watermark found at last, as writehuman.ai reports (source: writehuman.ai, https://writehuman.ai/blog/does-chatgpt-watermark-text). OpenAI’s explanation was less cinematic: those characters were artefacts, not a tracking system.

The other half of the story is a product that was built and shelved. writehuman.ai notes that OpenAI built watermarking technology two years earlier, rated it internally at about 99.9 percent effective, and then decided not to ship it (source: writehuman.ai, https://writehuman.ai/blog/does-chatgpt-watermark-text). A watermark that was never deployed cannot be read in your document.

So the anxiety is understandable and the evidence is thin. What follows is what you can check yourself, today, without trusting anyone’s marketing page.

How to check if ChatGPT added a watermark: four things you can actually verify

None of these methods will produce a watermark verdict, because there is no watermark to find. They will tell you whether your text carries hidden characters and whether it reads as machine-made, which is what actually causes trouble.

1. Inspect for invisible characters. Paste a sample into a plain-text editor or a Unicode inspector. Look for zero-width spaces, word joiners, and unusual directional marks. If you find them, strip them. They are noise, not a signature.
2. Read OpenAI’s current policy pages. Check what the model documentation says about output provenance and disclosure. If a watermark were live, it would be documented there first, not discovered in a forum thread.
3. Run the text through a detector that reports at sentence level. A single document score tells you little. A sentence-by-sentence view shows which lines carry the signal, so you can edit the right ones. Our own detector works this way, and you can read more about how the signals are built in How AI Detection Works.
4. Compare against your own writing. Put a paragraph you wrote before ChatGPT existed next to the flagged paragraph. The differences are usually rhythm, sentence length, and the density of stock transitions.

If you want a second opinion on the text itself rather than the watermark question, the AI watermark explainer covers what watermarks are and what they are not.

If you want to see which sentences carry the signal in your own draft, DeAIze scores the whole document and highlights each sentence by AI probability. The rewrite trial gives you 200 words after email confirmation, no card required — start here.

What our own tests actually show

We ran our detector on two sets of passages on 2026-09-28, using the local ensemble at a 30% flag threshold. This is our own measurement, not a third-party detector result, and it is not a false-positive rate.

The reference sample was 709 passages from our own guides that were not marked as generated. Selection was by metadata only, and human authorship was not independently verified. The AI score mean was 26%, the median 25%, and the range 0–69%. At or above the 30% threshold: 239 of 709. These are passage counts, not counts of independent documents.

The unedited model output was 12 passages generated from fixed prompts and not edited. AI score mean 33%, median 34%, range 4–46%.

Read that carefully. The two distributions overlap heavily. Unedited model output is not automatically flagged, and unedited human writing is not automatically cleared. Any tool that promises a clean watermark check is promising something the signal cannot deliver. Passages from the same guide are not independent samples, and passage construction targeted 140 words with shorter eligible tails.

What OpenAI says about watermarks, in practice

OpenAI signed the EU’s Code of Practice for general-purpose AI, as writehuman.ai reports (source: writehuman.ai, https://writehuman.ai/blog/does-chatgpt-watermark-text). Signing a code of practice is a governance commitment. It is not a product release, and it does not mean a watermark is embedded in the text you copied out of the chat window.

Watermarking text is genuinely hard. The scheme has to survive paraphrasing, translation, copy-paste through three different editors, and any user who deletes a clause. A method rated internally as highly effective in a controlled test can still fail the moment real editing starts. That is a plausible reason to shelve it, and it matches what users observe: no watermark check exists because no deployed watermark exists.

There is also a disclosure question that has nothing to do with watermarks. If your institution or client requires you to declare AI assistance, the watermark is irrelevant. You declare it. Our guide to AI content disclosure covers what to say and when.

The comparison: what each option actually gives you

The table below is ours. It compares what you get from each route when you are trying to work out whether ChatGPT text is traceable and what to do about a flag.

Option	What it reports	What it costs	Best for
OpenAI policy and model docs	Whether a watermark is documented as shipped	Free	Settling the watermark question
Unicode inspector	Hidden characters, zero-width marks	Free	Checking for stray invisible characters
DeAIze detector	Document score plus sentence-level AI probability	Configurable detection limits shown in the interface	Finding which sentences carry the signal
DeAIze rewrite	Suggested revised wording for your own draft	200-word trial after email confirmation; Starter $20.99 / 1,000 words, Standard $44.99 / 3,000 words, Pro $99.99 / 10,000 words, Max $199.99 / 30,000 words; credits never expire	Editing your draft before you submit it
WriteHuman	Humanization; its free plan covers three humanizations a month at 250 words each, no card required, per writehuman.ai	Free tier plus paid plans	Readers who want a humanizer comparison point

The WriteHuman figure comes from writehuman.ai (source: writehuman.ai, https://writehuman.ai/blog/does-chatgpt-watermark-text), which also notes the tool has led HumanizerBench for four cycles running, most recently in October 2026. That is their claim about their own benchmark, not our measurement.

The prices in the DeAIze row are ours and current. Detection limits are separate from the rewrite trial and are shown in the interface, so check there before you assume a volume.

When this is the wrong choice

Do not buy anything on this page if the watermark question is all you came for. The answer is free: no watermark is deployed, so no purchase resolves it.

Do not buy a humanizer if you are trying to hide the fact that someone else wrote your work. Rewriting is for your own draft. The ideas, data, and conclusions have to be yours, and a rewrite tool cannot make them yours. If you are inside an academic integrity process, the right move is to preserve your original files and process evidence, not to run the text through a tool. Our guide to what panels assess in an academic integrity hearing is the better starting point.

Do not buy a detection pass if your text is under roughly a couple of paragraphs. Short samples give detectors very little to work with, and the readings swing. There is a longer explanation in AI Detector Short Text Accuracy.

And do not buy a rewrite subscription if you are happy to edit by hand. A patient editing pass over your own draft does most of what a humanizer does, and it costs nothing.

A worked example

Here is an illustration, not a measurement.

Before: “The implementation of sustainable practices is a multifaceted endeavour that requires the collaboration of various stakeholders across multiple domains.”

After: “Cutting emissions at the plant took three departments and one uncomfortable budget meeting.”

The second version names a place, a number of departments, and a cost. It has a subject who did something. The first version has no actor and no event. That difference is what detectors pick up on, and it is also what readers pick up on. Our own tell-frequency measurement on 2026-09-28 came out at 1 per 100 words in both the reference sample and the unedited model output, which is a reminder that surface tells alone do not separate the two corpora.

Fix the substance first. If a sentence has no actor, give it one. If a paragraph makes a claim with no example, add the example. Then re-read for rhythm.

A checklist you can run in ten minutes

1. Paste a sample into a Unicode inspector and strip anything invisible.
2. Read the current OpenAI policy pages for any watermark statement.
3. Run the draft through a detector that shows sentence-level scores.
4. Rewrite only the highlighted sentences, keeping your claims, quantities, citations, negations, and qualifiers intact.
5. Compare the rewrite against the original line by line. Semantic similarity is a review aid, not proof that every meaning survived.
6. Keep the original file. If anyone questions the work, the draft history is your evidence.

If you are weighing whether paid editing help is worth it at all, AI Detector Tool Flagged You: When the Fix Is Worth Paying For walks through that decision without assuming you should buy.

---

Measurement note: figures in this article come from our own detector run on 2026-09-28 — n = 709 reference passages from guides not marked generated versus n = 12 unedited model passages, median AI score 25% and 34% respectively. Reference authorship was not independently verified; passages are not independent documents. Method: /methodology/.