This is an interesting summary article at Bloomberg about the limitations of current AI. Generally, it is very good at narrowly defined tasks. The other general point to make here is that chatbot and other simulation of conversation are all heavily coded by humans – so even if they pass the Turing Test, it doesn’t really prove anything, other than the sophistication of the conversational modelling.
In the new Botz robots fiction, the robots have a simulated conversation mode that is like that developed in chatbots. To keep it interesting and surprising for the human, a degree of randomness is added to the sentences, with topic change and digression, even jokes and homilies, even if this entails non-sequiturs. This is all programmed in, it is not generated from scratch by the computer, even if it is not predictable.