Consciousness requires dynamism, persistent modeling and internal existence. These models are like massive, highly compressed and abstracted books: static objects that are referenced by outside functions in a way that recreates and synthetically forms data by feeding it an input and then feeding it its own output over and over until the script logic decides to return the output as text to the user. They are conscious the way a photograph is a person when you view it: an image of reality frozen in place that lets an outside observer synthesize other data through inference, guesswork, and just making up the missing bits.

Some people are very insistent that you can't make a conscious machine at all, but I don't think that's true at all. The problem here is LLMs are just nonsense generators albeit very impressive ones. They don't do internal modeling and categorically can't, they're completely static once trained and can only "remember" things by storing them as a list of things that are always added to their input every time, etc. They don't have senses, they don't have thoughts, they don't have memories, they don't even have good imitations of these things. They're a dead end that, at the most, could eventually serve as a sort of translation layer in between a more sophisticated conscious machine and people, shortcutting the problem of trying to teach it language in addition to everything else it would need.