A question, answered
Why do chatbots sometimes miscount the letters in a word?
Partly because the model never sees the letters. Text is split into tokens before the model reads it, so a word arrives as one or a few multi-letter pieces, and counting the letters inside them means recalling how each piece is spelled, not looking at it. In two encodings from tiktoken, a widely used open-source tokenizer library (github.com), which we ran on 4 October 2026, " strawberry" with its leading space is a single token; without the space it is three pieces, such as "str", "aw" and "berry", none of them a single letter. Counting is also hard for these models in its own right. One 2024 study found that models could recognize the letters in a word but often miscounted them, especially letters that appear more than twice, like the three r's in "strawberry" (Fu et al., arxiv.org). Many newer models get this right, especially when they spell the word out first or reason step by step. Asking for the word one letter per line, then counting, usually helps. Try the exercise Spell it out first.
Read next: Spell it out first