I'm not sure negations are the crux of it, at least in terms of simple words like "not", which ChatGPT and other recent LLMs obviously use and understand. But

I'm not sure negations are the crux of it, at least in terms of simple words like "not", which ChatGPT and other recent LLMs obviously use and understand. But there is definitely some kind of blind spot there, where the model can't or won't explore "negative space". It wants to gravitate to common things - so if you try to get it to "think of something new" or producing something random that's not like these N other things, or tell you something that's uncommon (like naming any ordinary, non-famous person), it usually trips over itself.