Well, they can. But not all the way up to the top level of your program. But the longer you can hold off from your functions having side effects the more predictable and stable your codebase will be, with as an added benefit fewer bugs and less chance of runtime issues.

Pure functions often exist in practice and are useful for preventing many bugs. Sure, they may not be suitable for some situations but they can prevent a lot of foot guns.

Here's a Haskell example with all of the above:

  import System.Random (randomRIO)

  main :: IO ()
  main = do
    num <- randomRIO (1, 100)
    print $ pureFunction num

  pureFunction :: Int -> Int
  pureFunction x = x * x + 2 * x + 1