Does ChatGPT Watermark Its Text?
OpenAI built a ChatGPT text watermark and never shipped it. What is confirmed, what is myth, and how to check your text.
· 6 min read
Short version: OpenAI has confirmed it built a text watermarking system for ChatGPT, has said it is highly accurate, and has repeatedly declined to release it. Meanwhile, ChatGPT output does often contain invisible characters — just not for the reason most people assume.
Anthropic has now shipped one. Since August 2026, Claude embeds a machine-readable watermark in generated text worldwide under the EU AI Act. OpenAI signed the same Code of Practice, so ChatGPT's position may change — see how Claude's watermark works for what a shipped implementation looks like in practice.
What OpenAI has confirmed
OpenAI has publicly acknowledged an internal text-watermarking method and described it as effective against straightforward paraphrasing. It has also given its reasons for not shipping it: the technique degrades under translation and rewriting, it risks stigmatising non-native English speakers who legitimately use AI to edit their writing, and users surveyed said the feature would push them toward competitors that do not have it.
As of now there is no confirmed, publicly documented watermark active in ChatGPT output.
How text watermarking would actually work
Most people imagine invisible characters sprinkled through the text. The serious proposals do something quite different and much harder to defeat.
At each generation step the model produces a probability distribution over possible next tokens. A watermarking scheme uses a secret key to split the vocabulary into a pseudo-random "green" list and "red" list at every position, then nudges sampling toward green tokens. Any one word choice looks completely normal. Across a few hundred words, the proportion of green tokens is statistically far above chance, and whoever holds the key can measure it.
Nothing is inserted. Nothing is invisible. There is no character you can strip out, because the watermark is in which ordinary words were chosen. Rewriting the text in your own words destroys it; deleting characters does not.
So why does ChatGPT text contain invisible characters?
Because almost all text does. The usual sources are mundane:
- Non-breaking spaces (U+00A0) from the chat UI's HTML rendering, picked up when you copy
- Curly quotes and the … ellipsis character reproduced from the model's training data
- Zero-width joiners inside emoji sequences
- Narrow no-break spaces around punctuation in some locales
There have also been documented periods where specific model outputs carried consistent zero-width characters — reported for some Claude and GPT deployments — most plausibly artefacts of a tokenizer or rendering layer rather than deliberate marking. Either way they are trivially removable, which is exactly why no one designing a serious watermark would rely on them.
Check your own text
Do not take anyone's word for it, including this page. Paste your output into the ChatGPT watermark remover and open the Analyze tab. It lists every non-ASCII code point in the text with its unicode name and count. If there is a zero-width character in there, you will see it named; if there is not, you will see that too.
What this means practically
If you are worried about invisible characters — trivially checked, trivially removed, and this tool does both. If you are worried about a statistical watermark, no character-level cleaner can help, because there are no characters to clean; only substantive rewriting changes the token distribution. And if you are worried about AI detectors, note that they are not watermark readers at all: they are classifiers guessing from style, with well-documented false-positive rates against non-native English writers.
Related
Does Claude watermark its text? · The complete guide to invisible unicode characters
FAQ
+ − Does ChatGPT add a watermark to its text?
OpenAI has confirmed it built a text watermarking system but has not shipped it publicly, citing accuracy loss under paraphrasing and the risk of unfairly flagging non-native English speakers. There is no confirmed active watermark in ChatGPT output.
+ − Does ChatGPT add invisible characters?
Sometimes, but not as a watermark. Non-breaking spaces, curly quotes and zero-width joiners in emoji come through the chat interface's rendering. They are artefacts of copy-paste, and they are easy to detect and remove.
+ − How do I check my text for a ChatGPT watermark?
Paste it into an invisible character detector and review the analysis. It lists every non-ASCII code point with its unicode name and count, so any zero-width character present is named explicitly.
+ − Can a text watermark be removed?
Invisible-character marks can be stripped in one click. A statistical watermark embedded in word choice cannot be removed by cleaning characters — only substantial rewriting or paraphrasing changes the token distribution it relies on.
Clean your text now
Free, instant, and entirely in your browser — nothing is uploaded.
Keep reading
Does Claude Watermark Its Text?
Since August 2026 Claude embeds a watermark in all generated text, worldwide. What it is, what it proves, what removes it.
Invisible Unicode Characters Explained
A reference to invisible unicode — zero-width spaces, joiners, bidi marks and lookalike spaces — with code points and fixes.