A nondeterministic Turing machine accepts if at least one path accepts; it only rejects if all paths reject. This asymmetry makes it hard to "flip answers".
For example, suppose you have a nondeterministic Turing machine $M$ that has two paths for input $w$: one accepts, the other rejects. $M$ has at least one accepting path for $w$, so it accepts. Suppose we want to produce a machine that accepts exactly the inputs that $M$ rejects. The obvious first attempt is to take $M$ and make its accepting states reject, and its rejecting states accept. $M$ has one accepting path for $w$ and one rejecting path; this new machine $M'$ has one rejecting path and one accepting path. So it still accepts $w$, which it was supposed to reject!
A nondeterministic machine can't look at all its paths simultaneously and take action based on what all of those paths do. If you like, you can think of it as a form of parallelism where the threads are forbidden to communicate with each other. When all of the threads have finished the program must ask itself the following question: "Did at least one of my threads accept?" If the answer is yes, it is legally obliged to accept; if the answer is no, it is legally obliged to reject. It cannot do anything else.
When you simulate a nondeterministic machine $M$ using another one, $M'$, each path of $M'$ simulates one path of $M$ and sees only that path. It can't say, "If all those other paths rejected, I'll accept" because it can't see the other paths; it can only see itself. So all it could possibly say is things like, "If the path I simulated accepted, I'll reject" or "If the path I simulated accepted, I'll accept too". Then, at the end of the computation, the machine has to say, "If any of my paths accepted, I'll accept too", leading to the problem I described above. To invert the behaviour of $M$, each path of $M'$ needs to say, "If the path I simulated accepted, I reject; else, I accept" and, at the end of the computation, the machine needs to say, "If all of my paths accepted, I accept; else, I reject." This is because, if all of the simulator's paths accepted, that means all of $M$'s paths rejected, so $M$ rejected, so the simulator needs to accept. But the simulator isn't a valid nondeterministic Turing machine because it isn't using the legally mandated acceptance criterion. It can't do that.
The only way we know to figure out if a nondeterministic machine rejects its input is to try every possible path and verify that they all reject. After all, if even one of them accepted, the machine would accept the input. But trying every possible path is exponentially slower than trying just one.