Ask a chatbot how many R's are in the word strawberry, and for a long time the classic answer was a confident, wrong two. The real answer is three: strawberry. A system that can write working code and explain quantum physics tripping over a question a child gets right feels absurd — until you learn that the model never actually sees the word at all. This is a research-based explainer, drawn from the sources below, not a hands-on test of any chatbot.
The question became the internet's favourite chatbot test during 2024, passed around as a party trick that exposed something real. It grew so famous that researchers wrote papers about it, one literally titled "The Strawberry Problem", examining why character-level tasks fail. The joke worked because the failure was so specific: the model was not confused about fruit, and it was not bad at arithmetic. It was blind to the very thing being counted.
Before your prompt reaches the model, a separate piece of software called a tokenizer chops the text into chunks called tokens, and hands the model ID numbers for those chunks. On GPT-4o's encoding, strawberry becomes three tokens: "st", "raw" and "berry". The model does all its thinking over those ID numbers. The letters inside each chunk are invisible to it unless it happened to memorise them from training text. Asking it to count R's is like asking someone to count the letters in a word they have only ever heard spoken, never seen written down.
The chopping follows a method called byte-pair encoding. It starts from raw bytes and repeatedly merges the most frequent adjacent pairs into new tokens. Common words stay whole; rarer words get split into fragments the system has seen before. Nothing in predicting the next token forces the model to store, for each chunk, exactly which letters sit inside it.
Because it is a practical compromise. If every word needed its own token, the vocabulary would be enormous — every name, typo and technical term would need an entry. If every letter were its own token, every sentence would become far longer for the model to process, and cost more. Subword chunks sit in the middle: GPT-4o's encoding holds about 200,000 pieces, and a token works out to roughly three-quarters of a word on average. The packaging is efficient, but its shape constrains everything inside it.
Token blindness explains a whole family of odd failures. Spelling a word backwards goes wrong for the same reason. Long numbers get split into multi-digit chunks rather than clean single digits, which is one reason models fumble exact arithmetic on long figures. And text in languages that appear less often in training data tends to split into many more pieces, which means the same idea can cost more tokens — and more money — in one language than in another. These are not separate bugs. They are one design trade-off, showing up in different places.
The fix follows directly from the cause: give the model the letters. Ask it to spell the word out first, one letter at a time, and then count — spelled out with separators, strawberry becomes ten tokens instead of three, but now each letter is visible. Or ask it to reason step by step rather than blurting an answer. And for anything where exactness matters, such as counting characters in a password or checking a code, use a tool that counts for real rather than a model that guesses from memory. A chatbot is a brilliant meaning machine. It was never built to be a pair of eyes.
The strawberry test is worth remembering the next time an AI answer sounds certain. Confidence is not the same as seeing: a model can be fluent about something it has never actually looked at. When precision matters — letters, digits, exact counts — hand it the pieces, or hand the job to a tool that counts.
Related on Poptani Labs: Why AI sometimes makes things up — and why it sounds so sure, Why your chatbot now remembers you — AI memory, explained, How to run AI locally on your computer. More in this category: AI, Decoded.
Sources: AI/TLDR, on the strawberry problem; Toolbit, tokenization explained; Prompt20, tokens and tokenization