The newer GPT-o3 and GPT-o4 mini models appear to be embedding special character watermarks in generated text. However, removing these watermarks is relatively simple, making this seem more like a short-term measure than a long-term solution
Non breaking spaces and zero width spaces are not watermarks. They are used all the time in professional typesetting and it’s actually a good thing that GPT models can do that now.
But of course they can be a tell if you write your work email like a professional type setter but to be fair there are lot of other tells too in GPT outputs.
PS: In fact there are even keyboards (but they are rare) for example for German E1 extension (https://de.wikipedia.org/wiki/E1_(Tastaturbelegung)) that can even type those characters. They are used to prevent unwanted line-breaks, for example between numbers and there units or to allow for hyphenation in long words.
Yeah, if you look at the examples, they’re all added in places where you ideally don’t want the words to be split. This is just a typography thing they added to the output, not an intentional watermark.
Non breaking spaces and zero width spaces are not watermarks. They are used all the time in professional typesetting and it’s actually a good thing that GPT models can do that now.
But of course they can be a tell if you write your work email like a professional type setter but to be fair there are lot of other tells too in GPT outputs.
PS: In fact there are even keyboards (but they are rare) for example for German E1 extension (https://de.wikipedia.org/wiki/E1_(Tastaturbelegung)) that can even type those characters. They are used to prevent unwanted line-breaks, for example between numbers and there units or to allow for hyphenation in long words.
Yeah, if you look at the examples, they’re all added in places where you ideally don’t want the words to be split. This is just a typography thing they added to the output, not an intentional watermark.