Why does ChatGPT output contain U+202F?

U+202F is a narrow no-break space, and ChatGPT sometimes puts one where an ordinary space belongs. OpenAI's position is that it comes from how the model was trained, not from a watermark — and either way, finding one tells you nothing about who wrote the sentence.

The character is ordinary typography

Nobody invented it for AI. French punctuation puts a narrow no-break space before ; : ! and ? and inside the quotation marks « ». French number formatting uses it as the thousands separator, so 1234567.89 is set as 1 234 567,89. Mongolian uses it between a word and certain suffixes. Word processors, typesetting systems and translation tools all emit it.

A model trained on a large amount of professionally typeset text picks that habit up the way it picks up any other. That is the explanation OpenAI has given, and it fits: the character shows up inconsistently, it varies between models, and a real watermark built this way would be defeated by one find-and-replace.

Where the watermark story came from

In April 2025 the team at Rumi tested OpenAI's then-new o3 and o4-mini models, found narrow no-break spaces scattered through longer outputs, and asked whether they were watermarks. OpenAI said they were not — a quirk of large-scale reinforcement learning — and within a few days Rumi retested and the characters had stopped appearing.

They came back. Developers filed the same complaint about GPT-5 output on OpenAI's own forum in late 2025, where in some macOS apps the character renders as a visibly squeezed gap instead of a space.

What nobody has published is a rate: how often it appears, in which models, under which prompts. Any tool that gives you a confidence score based on this character is inventing it.

What to actually do with one

If it is breaking your rendering or your diff, replace it with a normal space. Paste the text into the hidden character checker and it will list every unusual codepoint, U+202F included, and give you the text back clean.

If you are trying to work out whether someone used ChatGPT, this cannot tell you. It survives copy-paste, which means it also survives a human quoting ChatGPT, and it arrives in text from Word, a CMS, a PDF export or any French-language document. Why that inference fails.