Words are generated word by word. It’s reading the entire prompt and what it replied so far to generate new words. So yeah it can recognize its own mistakes while writing. It just wasn’t trained for that so it usually doesn’t do it, but you can encourage it to do it by giving custom instructions telling it to second guess itself.