Why does an AI chatbot that can spell out the word 'strawberry' letter by letter still so often miscount how many r's it contains?

Answer: It reads text in chunks rather than letters, and the tally it rebuilds from them slips when a letter repeats

It learned from web text where people miscount the word themselves, so it copies their mistakes — Wrong. People rarely miscount 'strawberry' in writing, so there is no flood of wrong answers to copy. The error comes from how the model processes text, chunk by chunk, not from imitating human mistakes.

It reads text in chunks rather than letters, and the tally it rebuilds from them slips when a letter repeats ✓ — Correct! A chatbot's language model does not see letters; it sees text cut into chunks called tokens, so 'strawberry' arrives as two or three pieces. To count r's it has to rebuild the spelling from those chunks and keep a running tally, and the tally is the weak link: a 2024 study of letter-counting errors found that models could name the letters in a word but failed most often when a letter appeared more than once, while how common the word was made no significant difference.

It stops searching at the first r it finds, so the second and third r never get counted at all — Wrong. Chatbots do not give up after the first hit; asked directly, they will usually say the word has more than one r. The miscount happens while tallying repeated letters, not because the search stops early.