Why should they be? The token is at the word level more or less, so the model can't see the letters to substitute.

When we get character-level models this should be more viable.