AI Watermark: What It Is and How to Fix Flagged Text

An AI watermark is a hidden statistical or invisible-character signal that a model can embed in its output. Most of the time, the flag on your draft is not a watermark at all — it is a detector scoring your text. The fix depends on which one you have, and they are not the same problem.

What an AI watermark actually is (and what it is not)

There are two different things people call a watermark.

Statistical watermarks. A model nudges word choices during generation using a secret key. The text reads normally to you, but a matching detector can test whether the pattern fits. OpenAI built exactly this technology, rated it internally at about 99.9 percent effective, and then decided not to ship it. That decision is now under real pressure; OpenAI signed the EU’s Code of Practice for general-purpose AI in July 2025, according to writehuman.ai (https://writehuman.ai/blog/does-chatgpt-watermark-text).

Invisible character watermarks. These are Unicode characters you cannot see — zero-width spaces, unusual separators — inserted into text. In the spring of 2025, users spotted unusual invisible Unicode spacing characters in ChatGPT output and declared the secret ChatGPT watermark found at last. OpenAI’s explanation was less cinematic, per the same writehuman.ai piece (https://writehuman.ai/blog/does-chatgpt-watermark-text).

Here is the part most articles skip. Neither of these is what usually gets your essay or article flagged. Detectors like GPTZero, Turnitin, Copyleaks, Originality.ai and Sapling score the statistical properties of your text — perplexity, burstiness, sentence rhythm — and return a probability estimate. That is not proof of authorship, and it is not a watermark.

If you want the mechanics, our guide on how AI detection works walks through the signals. The short version: a high score means your text looks statistically typical of model output. It does not mean a hidden mark was found.

Why does my human writing get flagged as AI?

Because the signals overlap. Human writing and model output share features: common transitions, uniform sentence length, predictable vocabulary. When your draft has those features — because you write cleanly, or because English is not your first language, or because you edited heavily with a tool — the detector has no way to separate you from a model.

The archived 2026-09-28 DeAIze local detector observation used a 30% flag threshold. Its reference group contained 709 passages from 61 guides selected by metadata; human authorship was not independently verified. Mean score: 26%; median: 25%; range: 0–69%. At or above the threshold: 239 passages. These are local readings, not an independently measured human false-positive rate.

The same observation included 12 unedited DeepSeek passages from fixed prompts, with a mean score of 33%, median 34% and range 4–46%. Their readings overlap the reference range. Because reference authorship was not independently verified and per-passage records were not retained, this overlap cannot establish detector accuracy or settle authorship.

If you want to see which sentences in your draft carry the signal rather than just a document score, the AI detector highlights each line by AI probability so you can edit the specific ones that matter.

The fix depends on which problem you have

Work through this table before you touch anything.

What you have	How to confirm it	What to do
Statistical watermark (rare, model-specific)	You pasted raw model output with no edits and a matching detector exists	Rewrite the flagged sentences yourself; a watermark does not survive genuine rewriting
Invisible Unicode characters	A text editor shows hidden characters, or the text behaves oddly when copied	Strip non-printing characters, then re-read for meaning
Detector flag on your own writing	A detector returns a probability and highlights sentences	Edit the flagged sentences for rhythm and specificity; keep your meaning
Detector flag on your AI-assisted draft	Same as above, but the structure is also generic	Restructure before you rewrite — see humanize AI text

Most people reading this are in rows three or four. The watermark rows are real but far rarer than the internet suggests.

A worked example

Here is an illustration, not a real draft.

Before (reads machine-made): “It is important to consider that artificial intelligence has fundamentally transformed the landscape of content creation, enabling unprecedented opportunities for efficiency and scalability.”

After (keeps the claim, loses the pattern): “AI tools made it faster to produce content at scale. That changed who could publish, and how much.”

The claim survived. The filler did not. This is the whole method: cut the abstract nouns, break the uniform rhythm, and say the specific thing.

When this is the wrong choice

Do not humanize text that is not yours. If you generated a draft and the ideas, data and conclusions are not your own, rewriting the surface does not make it your work. That is misrepresentation, and it is what academic integrity panels actually assess — see academic integrity hearing AI for what they look at.

Do not humanize to defeat a detector while submitting work you did not do. And do not assume humanizing is necessary at all. If your draft is your own writing and a detector flagged it, the honest move is often to document your process, not to rewrite. Our piece on false positives covers that path.

Also skip the paid tools if your text is short. Some free plans cover a few humanizations a month at limited word counts — writehuman.ai’s own page lists three humanizations a month at 250 words each on its free plan (https://writehuman.ai/blog/does-chatgpt-watermark-text). If your draft is longer than that, a free tier will not finish the job.

How DeAIze handles it

DeAIze (https://deaize.com) is a web app with two halves: an AI detector and an AI humanizer.

Detection is sentence level. 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.

Humanizing runs a closed loop: score the text, rewrite the flagged sentences, apply an automated semantic-similarity check, then re-score and report before and after. Compare each suggestion with the original; similarity alone cannot establish that facts and qualifiers survived. The retained six-draft rewrite observation showed mixed score changes across models and does not substantiate a current customer average, a 60%–71% reduction or a 92%-to-4% expected result. It included no commercial detector evaluation.

A DeAIze reading does not predict an official Turnitin result. If you obtain an actual third-party report, record its provider, date and returned output separately. Uploads cover .txt, .docx and .pdf, and rewriting a Word document preserves fonts, styles and tables. Detection and rewriting are English-first and cover English plus nine other languages. Text is processed for the current task only, never used for training and never resold. The free rewrite trial gives you 200 words after email confirmation, no credit card; detection has separate limits.

Detector scores are probability estimates, not proof of authorship. No tool can guarantee a particular score on any detector.

What to do next

1. Confirm what you actually have. Paste the raw text into a detector and read the sentence-level highlights, not just the document score.
2. If you see hidden characters, strip them. If you don’t, stop looking for a watermark.
3. Edit only the flagged sentences. Keep the ones that already read like you.
4. Re-score. If the flagged lines dropped and the meaning held, you are done.
5. If it is your own writing and it still flags, document your process rather than rewriting endlessly.

The uncomfortable truth: an AI watermark is mostly a red herring for the person whose essay got flagged. The real work is editing the sentences that carry the signal, and being honest about which parts of the draft were yours.

---

Measurement note: the archived 2026-09-28 local observation contains 709 reference passages from 61 metadata-selected guides and 12 unedited DeepSeek passages. Median scores were 25% and 34%. Reference authorship was not independently verified; these figures do not establish a human false-positive rate. Method: /methodology/. See the archived observation.