How would we know it was correct?

If you feed an AI nonsense in its training data, it will generate nonsense