Does ChatGPT Watermark Its Text? What We Actually Know
· 5 min read · GhostVerb team
It's one of the most searched questions about AI writing: if you use ChatGPT, is there something hidden in the text that marks it as AI? The short answer is not in any confirmed, systematic way, but you can't assume your text is clean either. Here's what's actually known.
What OpenAI has said
OpenAI has said publicly that it built a working text watermarking method. In 2024 it explained that it was holding back from rolling it out broadly. One concern was that watermarks can be weakened by rewriting or translating text; another was the effect on people who use AI as a writing aid, including non-native English speakers. OpenAI has also discussed other approaches, such as attaching metadata to content. Policies can change, so treat any "ChatGPT definitely doesn't watermark" claim as a snapshot in time.
What users have found
In 2025, users noticed that text from some newer ChatGPT models contained narrow no-break spaces (U+202F), characters that look exactly like normal spaces. Posts calling it a secret watermark spread quickly. OpenAI said it was an unintended quirk of the models, not a watermark.
Whatever the cause, the practical point stands: AI text can contain characters you can't see, and some tools and readers notice them. We explain the common ones in our guide to zero-width spaces in ChatGPT text.
How Gemini is different
Google is open about watermarking. Text generated by Gemini in Google's apps carries SynthID, a statistical pattern in the word choices rather than hidden characters. It's invisible, survives copy and paste, and can only be checked by someone holding Google's key.
And AI detectors?
Most "AI detectors" don't look for watermarks at all. They guess from style: predictable word choices, even sentence lengths, and favourite AI words like "delve" and "tapestry". They're known to be unreliable in both directions.
How to check your own text
- Paste it into the free hidden character checker. It shows every invisible or look-alike character, and nothing leaves your browser.
- Remove anything it finds with one click.
- For wording-level watermarks like SynthID, there's nothing to see. Only a full rewrite removes them.
The safest approach
If you want text that carries no AI fingerprints of either kind, it needs clean characters and fresh wording. That's what GhostVerb does: it strips hidden characters and rewrites every sentence in your own voice, keeping your meaning, facts and tone.