Because LLMs dont have a textual representation of any text they consume. Its just vectors to them. Which is why they are so good at ignoring typos, the vector distance is so small it makes no difference to them.