What you call the « next word » is actually the next word it has already seen in a similar context in the training set.
The novelty does no go beyond the variability of the training data.
And so deep learning cannot generate futures that are both surprising and general.
In the context of conversations, no doubt that it can generate realistic answers. They will just be regurgitations of the training data. They might seem novel to you because you haven’t experienced all the discussions of the training set. But they won’t be novel to humankind and won’t project it into an actual future.
The only reason would be hitting the hardcoded input length limit.