Humans also would probably be very likely to guess 2 r's if they had never seen any written words or had the word spelled out to them as individual letters before, which is kind of close to how lanugage models treat it, despite being a textual interface.