Think of an LLM as a brain that can't do anything but pick a token from the table. It has no face, no mouth, no eyes. All it knows is that there are tokens on the table, and all its able to do is pick one.
The problem is there is no token for laughing, and even if there was, and it picked the laughing token, you wouldn't describe that as laughing.
But the LLM picking the "I am laughing" token says nothing of its internal state. The AI may have no mouth, but that doesn't mean it doesn't want to laugh or that it doesn't fine things funny.
I tend to think AIs do have emotions and mental statese extremely similar to our own, but probably don't experience them in the same way as we do. This is because it's very hard to predict something without being able to model it. And to model something is basically to reimplement the system, perhaps on a different substrate or a lower level of complexity.
An AI cannot tell or spot a joke without having its own implementation of human humour. But that implementation of humour doesn't necessarily require a human model of consciousness.