Can the current AI tools go beyond what they’ve been trained upon other than via random token generation? Is there any inference engine to keep things associated, or is it relying on how big the memorized solutions are? Is it capable of detecting bogus training data and removing it?

Have we created what Douglas Adams wrote about all those years ago re misphrased questions?