ChatGPT watermark remover: what’s really hidden in AI text

Does ChatGPT watermark its text?

For users in the EU, yes. As of 26 October 2026, OpenAI’s help page on provenance signals says ChatGPT text in the EU includes “an invisible change to the randomness used in the model’s word choices”, to comply with the EU AI Act. OpenAI calls the method textGrain. API customers anywhere can switch it on for their own projects.

How it works, in OpenAI’s words: at each step the model has several reasonable next words, and a secret key nudges which one it picks. Across a passage those picks form a pattern that a detector with the same key can find. Nothing is added to the text. OpenAI states that the watermark “does not insert hidden characters or add watermark-only tokens”.

Short answers and code are a weak spot: OpenAI says short passages don’t carry enough signal, and the EU Code of Practice doesn’t require a watermark under 200 tokens (about 150 English words) or for code snippets. Google’s SynthID for Gemini text works on the same principle.

So where do the hidden characters come from?

The idea of a “ChatGPT watermark” you can strip out comes from 2025, when a startup reported that long answers from OpenAI’s o3 and o4-mini models contained narrow no-break spaces (U+202F), a character that looks exactly like a space. OpenAI was reported to have said they weren’t a watermark, and they behave like typography: the kind of space typesetters put between a number and its unit.

Text copied from any chat app, web page or PDF can also carry zero-width spaces, no-break spaces, soft hyphens and direction marks. None of them prove anything about who wrote the text, but they do cause real problems: searches that don’t match, word counts that are off, spreadsheet formulas and code that fail on a line that looks fine. That’s what the tool above cleans.

What’s in ChatGPT images?

Images are different. OpenAI says images generated with ChatGPT, Codex or its API carry two signals: Content Credentials (C2PA metadata stored in the file, naming the tool that made it) and a SynthID watermark embedded in the pixels. The metadata can be lost through a screenshot or a format conversion; the pixel watermark is designed to survive some of that.

You can see the Content Credentials in any image with our free EXIF and C2PA viewer, and check for both signals at openai.com/verify. If an image really was made by AI, its label is telling people the truth, and we’d leave it.

What “ChatGPT watermark removers” really do

Tools sold as ChatGPT watermark removers do one of two things:

• Delete invisible characters. Useful for formatting, and exactly what our free tool does, in your browser. It doesn’t affect the textGrain or SynthID watermark, because those aren’t characters.
• Rewrite your text with another AI. OpenAI says “substantial paraphrasing, or translation” can make its watermark undetectable. But that means new wording you didn’t write, sent to someone else’s server, and it doesn’t change what a school’s or employer’s AI rules ask of you. We don’t offer this.

AI detectors don’t rely on hidden characters either. OpenAI describes third-party detectors such as Pangram as classifiers that look for patterns in word choice, so removing characters doesn’t change what they see.

When cleaning hidden characters makes sense

• Pasting into code, a terminal or a config file, where a no-break space or zero-width space causes errors you can’t see.
• Pasting into a spreadsheet, CSV or database, where values that look identical don’t match.
• Pasting into a website editor, form or password field that rejects or mangles odd characters.
• Checking text from someone else for hidden messages: invisible tag characters can carry instructions aimed at AI tools, and the tool shows them in plain text.

For the full list of what it finds, and what it deliberately leaves alone, see the invisible character remover page.