Does ChatGPT Watermark AI Text? The Honest Answer

No, ChatGPT does not watermark its output in any way a detector reads. The uncomfortable part: OpenAI shelved watermarking, and no watermark is why flagged writers have no clean proof of origin. What you actually face is a statistical flag, not a hidden tag — and that flag responds to editing.

Does ChatGPT watermark AI text — or is this a rumour?

OpenAI built watermarking technology and chose not to ship it. According to writehuman.ai’s report on the topic, OpenAI rated the approach at about 99.9 percent effective internally and then held it back, a decision that did not survive Europe’s new rules (source: writehuman.ai, https://writehuman.ai/blog/does-chatgpt-watermark-text). The same report notes that in the spring of 2025, users spotted unusual invisible Unicode spacing characters in ChatGPT output and declared the secret watermark found; OpenAI’s explanation was less cinematic (source: writehuman.ai, https://writehuman.ai/blog/does-chatgpt-watermark-text).

That history matters because it explains the gap you are standing in. A watermark would be a deliberate, verifiable mark. A detector score is not. It is a model’s estimate of how statistically likely a passage is to have come from a language model. Different vendors define that number differently — a model score, a share of qualifying prose and a similarity percentage are not interchangeable, and none of them proves authorship.

So the honest answer to “does ChatGPT watermark ai text” is no, and that is not good news. It means you cannot point at a hidden tag and clear your name. You can only look at what the detector actually measured and fix the writing that triggered it.

What the detector is actually scoring

Detectors read surface statistics: how predictable each word is, how much sentence rhythm varies, how often certain phrases recur. Machine-made drafts tend to be smooth and even. Human drafts wobble.

We ran our own test on this, and the numbers are worth reading carefully. On 2026-09-28 we scored two corpora with the DeAIze detector (local ensemble, lite tier) at a 30% flag threshold. A reference sample of 709 passages from DeAIze guides not marked generated — selected by metadata only, so human authorship was not independently verified — scored a mean of 26% and a median of 25%, with a range of 0 to 69%. Of those 709 passages, 239 landed at or above the 30% threshold. Those are passage counts, not counts of independent documents, and not a verified human false-positive rate.

Read that again, because it is the least comfortable number on this page. Even in writing we produced ourselves, roughly a third of passages crossed a threshold that would look like an accusation in a classroom or an inbox. A flag is a signal about style, not a verdict about who typed the words.

The unedited model output told the opposite story: 12 passages generated from fixed prompts and left unedited scored a mean of 33% and a median of 34%, range 4 to 46%. The spread overlaps. No single reading separates the two groups cleanly.

If you want to see which sentences in your own draft carry the signal, paste it into DeAIze — detection highlights every sentence by AI probability, and the rewrite trial gives you 200 words after email confirmation with no credit card. Compare plans at /pricing/.

Comparison: what each option actually gives you

Option	What it is built for	Limits and trade-offs	Best when
Writing it yourself, then editing	Full authorship and control	Slowest path; you still have to revise	The work must be defensibly yours
ChatGPT or another model	Generating and drafting	No watermark; output reads smooth and even, which detectors notice	You need a starting point, not a final draft
Detection tools (DeAIze, GPTZero, Turnitin and others)	Estimating how machine-like a passage reads	Score definitions differ; no score proves authorship	You want to see which sentences carry the signal
Rewriting tools (WriteHuman, DeAIze humanizer, others)	Suggesting revised wording for your own draft	You must verify claims, quantities, citations, negations and qualifiers yourself	You have a draft worth keeping and need it to sound like you
Doing nothing	—	The flag stays on the record	Never, if the flag has consequences

One figure from the third-party side, for calibration: WriteHuman’s public monthly benchmark, HumanizerBench, has been led for four cycles running by one tool, most recently in October 2026, and its free plan covers three humanizations a month at 250 words each with no card required (source: writehuman.ai, https://writehuman.ai/blog/does-chatgpt-watermark-text). That is a vendor’s own benchmark, so treat it as a claim to test, not a fact to trust.

The fix: a sentence-level editing pass

This is the part that actually changes your outcome. Work through these in order.

1. Scan first, edit second. Run your draft through a detector that shows sentence-level scores. In DeAIze, the whole document gets a score and each sentence is highlighted by AI probability, so you can see which lines carry the signal instead of guessing.
2. Rewrite the flagged sentences by hand. Not the whole document. The flagged lines are usually the ones with even rhythm, stacked transitions and generic verbs. Break the rhythm. Cut the throat-clearing opener. Put the subject at the front.
3. Add what only you know. A date, a place, a number you measured, the reason you chose one method over another. Detectors cannot model specifics you invented, because they were not in the training distribution of generic prose.
4. Check meaning, not just wording. Go through every claim, quantity, citation, negation and qualifier against your original. Semantic similarity flags are a review aid, not proof that every meaning survived.
5. Re-scan and stop when it is good enough. Detector scores move between runs and between tools. Chasing a perfect number wastes your afternoon. Get it below the threshold your institution or client uses, then stop.

Here is a labelled illustration of step two.

Before (flagged): “Effective communication plays a crucial role in the success of any organisation, and therefore leaders must prioritise clear messaging at all levels.”

After (edited): “Leaders who write badly get ignored. In our last project, the weekly update ran to nine paragraphs and nobody read past the second.”

The second version is shorter, contains a specific detail, and varies sentence length. That is the whole method. If you want the full workflow, the guide at /humanize-ai-text/ walks through it section by section.

When this is the wrong choice

Do not buy a humanizer if any of these describe you.

• You are trying to pass off someone else’s work as your own. That is a different problem, and no tool fixes it. The ideas, data and conclusions have to be yours.
• You are in an active misconduct dispute. Preserve the original, keep version history, and follow the appeals process. A rewritten document is not evidence. See the guide on academic integrity hearings at /blog/academic-integrity-hearing-ai/ for what panels actually assess.
• Your institution bans AI assistance outright. Check the policy first. A tool that improves your prose does not change what the policy permits.
• Your text is already fine. A single low-probability sentence is not a crisis. Save the money.
• You want a score guarantee on another vendor’s detector. Nobody can give you that, including us. DeAIze does not guarantee any particular score elsewhere, and a vendor API was not used to produce the figures above — those are our own measurements.

What to do next, in order

First, stop looking for a watermark. There is not one, and the search will not clear you. Second, find out what was actually measured — which tool, which threshold, which sentences. Ask for the report if you do not have it. Third, run your own scan so you are not arguing from ignorance. Fourth, edit the flagged sentences yourself and re-scan. Fifth, if the flag has consequences, keep your drafts and your version history, and use the disclosure route your institution or client provides rather than improvising.

That is the honest shape of this problem. OpenAI does not tag your output. A model estimates how machine-like it reads. The estimate is noisy on both sides — our own reference passages crossed a 30% threshold 239 times out of 709. And the only lever you control is the writing itself.

---

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/.