I think these models struggle with this because they don't process text as individual characters, but rather as tokens that often contain parts of a word. So the model never sees the actual characters within a token, and can only infer the contents of a token from the training data itself if the training data contains more information about it. It can get it right, but this depends on how much it can infer from training data and context. It's probably a bit like trying to infer what an English word sounds like when you've only heard 10% of the dictionary spoken aloud and knowing what it sounds like isn't actually that important to you.
More info can be found here: https://platform.openai.com/tokenizer
