At its most basic level, GenAI takes a prompt and assigns each word (technically tokens) a meaning and a ranking of the attention/relative importance. From there, this analysis is bounced against giant a correlation matrix of the probability of one word following another word. The model then generates a response on a word by word (token by token) basis, where the next word is based on the previous word.
This is a gross oversimplification. But it’s largely true.
Ergo, you’re right. The bot doesn’t know what it’s doing.
Which makes the whole AI "conscienceness" issue moot.
I think what is taken for AI being sentient is a hacker down in the bowels of the CIA.