Does ChatGPT actually watermark its text?
OpenAI has never shipped a confirmed, deliberate text watermark in ChatGPT. The company has said publicly that it built and tested a text-watermarking system and chose not to release it, largely because it is easy to defeat and would disadvantage non-native English speakers who paste output through translation tools.
That is the honest answer to the question as asked. But it is not the answer to the question most people actually mean, which is: why does my pasted ChatGPT text keep tripping checkers, and why does it behave oddly in Word?
The practical reality is that ChatGPT output very often does contain invisible characters, whether or not anyone intended them as a mark. Narrow no-break spaces appear around numbers and units. Non-breaking spaces show up where a plain space belongs. Zero-width characters get introduced during formatting and markdown conversion. These are real characters sitting in your document. They survive copy and paste, and they are trivially visible to anyone who runs a hex view or a Unicode checker over the file.
So rather than argue about intent, this tool just shows you what is in your specific document. If it is clean, it says so.
What we typically find in ChatGPT output
The most common finding by a wide margin is U+202F, the narrow no-break space. ChatGPT uses it in places a person would type an ordinary space: before units, inside numbers, around certain punctuation. On its own it is harmless, but it is a strong stylistic fingerprint because almost nobody types it deliberately.
Next most common are typographic characters that are not invisible but are equally telling: the em-dash (—), curly quotation marks, and the single-character ellipsis (…). These are not watermarks in any meaningful sense. They are just what the model produces, and they are also the fastest way for a human reader to spot pasted output, which is why this tool offers to normalise them.
Genuine zero-width characters (U+200B, U+2060, U+FEFF) turn up less often in raw ChatGPT output but frequently once text has passed through a web page, a Google Doc, or an email client. If your text has been anywhere before it got to you, they are worth checking for.
What removing these does and does not achieve
Removing invisible characters removes invisible characters. That is a complete and verifiable result: run the check again afterwards and the count is zero.
It does not change what your writing sounds like, and most AI detectors used in education are working on writing style, not on hidden characters. If a detector flagged your text, stripping a few narrow spaces will most likely not change its verdict. We would rather tell you that up front than take your money for a result you will not get.
Where this genuinely helps: knowing what is in a file before you submit it, cleaning up text that picked up junk in transit, and fixing documents that behave strangely because of stray control characters.