Does ChatGPT Watermark Its Output? A Straight Answer

Search this question and you will find confident answers in both directions. One set of posts insists OpenAI embeds a secret signature in every response so it can trace your essay back to you. Another insists the whole thing is a myth and there is nothing in your text at all.

Both are wrong, in the way that answers tend to be wrong when the question contains two different things wearing the same name.

The short answer

No, not in the sense people mean. There is no known hidden identifier in ChatGPT's output that ties a passage back to your account, your session, or your prompt. OpenAI has researched text watermarking publicly and has not shipped it as a deliberate signature on general output.

But two real things do travel with your text, and conflating them with a secret signature is what generates all the confusion.

Trace one: the characters

Copy a response out of the chat interface and you take formatting characters with it. Not because anything is marking you — because the interface is a web page, and flattening rendered HTML into plain text leaves behind the characters that were holding the layout together.

Zero-width spaces. Non-breaking spaces. Word joiners. Byte-order marks. Occasionally DOM attributes like data-start that ride straight into a CMS. Full breakdown in ChatGPT hidden characters.

These are real, they are in your document right now if you pasted from a chat window, and they persist through every save until something removes them. What they are not is evidence of anything. Text copied from a PDF, a Wikipedia page, or a colleague's email carries the same characters. Nobody is scanning for zero-width spaces and concluding you used AI — the false-positive rate would be catastrophic, because half the text on the internet would trip it.

They matter for a duller reason: they break things. Find-and-replace misses words. Word counts disagree. Keywords silently stop matching in a title tag. WordPress blocks scramble on save. Citation managers fail to link a reference. Each symptom costs an afternoon and none of them points at the cause.

Trace two: the style

This is the one that actually affects how your work is received, and it is not hidden at all.

Language models write distinctively. Not badly — distinctively. ChatGPT's defaults are recognisable enough that people who read a lot of it can spot a paragraph in seconds:

• Stock connectives carrying weight they have not earned: "furthermore", "moreover", "it's important to note", "in today's fast-paced world"
• Em dashes at roughly three times the rate anyone types them by hand
• Sentence lengths clustered in a narrow band, with none of the spikes and dips human writing has
• Compulsive both-sides hedging that never commits to a position
• Everything arranged in groups of three

There is no signature to find here because the writing is the signature. It cannot be removed by deleting characters. It requires the sentences to be rewritten.

Why the myth persists

Because the two traces get merged, and because "secret watermark" is a better story than "the copy button flattens HTML."

There is a genuine version of the deliberate-watermark story, just not for ChatGPT. Google's SynthID does exactly what people imagine — a real, intentional statistical watermark applied to Gemini output, sitting in the model's word choices rather than in hidden characters. It is worth understanding on its own terms: what SynthID actually is. Its existence is probably why the belief spread to every other model.

What this means for you

The practical version is short. The characters are a hygiene problem with a mechanical fix. The style is a quality problem with an editorial fix. Neither is a conspiracy, and the second one is the one that costs you something.

Doing both in one pass

On the PassMyEssay homepage, paste your draft and hit Clean:

• Every invisible character is stripped in a single deterministic pass — a lookup, no model, no promises needed.
• The stock phrasing is rewritten into prose with actual rhythm, holding your numbers, dates, quotes, and citations exactly where they were.

If you would rather see before you touch anything, switch to AI Phrase Detector. It flags the hidden characters where they sit and highlights the sentences that read as machine-written, tells you why each was flagged, and changes nothing until you do.

One honest limit while we are being precise about claims: we will not promise you a specific score on a specific detector. Detectors update constantly and their output is probabilistic, and anyone quoting you a number is quoting something they cannot verify. What we will promise is text that comes out clean in the byte stream and reads like a person wrote it.

Related: why AI writing sounds robotic, zero-width spaces explained, how AI detectors work.

Try PassMyEssay free → Strip what the copy button added, fix what the model wrote.