What Does the ChatGPT Watermark Look Like?
Like nothing at all, which is exactly why people find it unsettling. Here is what is actually in the text, character by character.
See it in your own text
0 chars
Your text is processed in memory and never stored. See the full list of characters this removes.
The short answer
You cannot see it. Every character involved either renders as zero pixels, or renders as something visually identical to an ordinary character. If you could see it, it would not have gone unnoticed for as long as it did.
What you can do is count it. Paste text into the box above and every flagged codepoint is listed by category.
The characters, side by side
Each row shows a normal character, the lookalike, and the codepoint you would find if you inspected the bytes.
| What you expect | What can be there | Codepoint | Difference |
|---|---|---|---|
| Ordinary space | Narrow no-break space | U+202F | Slightly narrower, never breaks a line |
| Ordinary space | Non-breaking space | U+00A0 | Identical width, never breaks a line |
| Nothing | Zero-width space | U+200B | Renders as nothing, splits a word for search |
| Nothing | Byte-order mark | U+FEFF | Renders as nothing, often at the start of a paste |
| Latin o | Cyrillic о | U+043E | Pixel-identical in most fonts |
| Latin a | Cyrillic а | U+0430 | Pixel-identical in most fonts |
| Latin e | Cyrillic е | U+0435 | Pixel-identical in most fonts |
| Bold text | Mathematical bold | U+1D400+ | Looks bold, is not markup, breaks search |
How to see them yourself
The reliable way is to look at the bytes. Save the text to a file and run it through a hex dump, or paste it into any Unicode inspector that lists codepoints. In VS Code, the "Render control characters" setting and the unusual-characters highlight will show most of them.
The unreliable way is looking at the screen, which is the whole problem.
What the 2025 U+202F story was really about
In April 2025, users noticed that OpenAI's o3 and o4-mini models were producing the narrow no-break space where an ordinary space belonged. Because it appeared consistently and invisibly, the natural conclusion was a deliberate watermark.
OpenAI described it as a quirk of large-scale reinforcement learning rather than a marking system, and the behaviour stopped appearing within days. The wider habit of AI output carrying non-standard characters has not stopped, which is why a cleaner is still worth having.
FAQ
Frequently asked questions
- Can I see the watermark by changing the font?
- Not reliably. A monospace font makes some spacing differences visible, but zero-width characters render as nothing in every font, and homoglyphs are identical in most. Counting the codepoints is the only dependable method.
- Does it show up if I paste into Notepad?
- Plain-text editors strip formatting but keep the characters, so no. Notepad++ with "Show All Characters" enabled will reveal some of them.
- Is the narrow no-break space always a sign of AI?
- No. It appears legitimately in French typography and in text copied from many websites and word processors. It is a signal worth checking, not proof of anything.
Last reviewed .