Ask a chatbot how many times the letter “r” appears in “strawberry.” For a while, many popular models confidently said two. The right answer is three. It became a running joke: AI chatbots miscount letters in a task a first-grader handles, even while they summarize legal documents and write sonnets. The mistake is not random, and it doesn’t mean the system is stupid. It tells us something real about how these large language models see text.
Why AI chatbots miscount letters: they don’t see letters
When you type a word, you see characters. A language model doesn’t. Before text reaches the neural network, a step called tokenization chops it into pieces called tokens. A token might be a whole common word, a chunk of a rarer one, a punctuation mark or a few characters.
So “strawberry” might arrive as two or three chunks, say “str,” “aw” and “berry,” depending on the model’s vocabulary. Each chunk is mapped to a number, and the network works with those numbers. The individual letters inside each chunk aren’t directly visible. Asking it to count letters is a little like asking you to count the number of times a certain brushstroke appears in a painting you’ve only seen from across the room.
Why tokenization exists
It helps to see why the design looks this way before judging it.
Models could, in principle, read one character at a time. But sequences would be much longer, which makes everything slower and more expensive, and it makes it harder for the model to see big-picture patterns. Tokens are a compromise: shorter sequences and a vocabulary of frequent chunks, at the cost of blurry spelling. That blur is the reason AI chatbots miscount letters while handling meaning so well.
So how does it ever get spelling right?
Models learn about letters indirectly, which is why AI chatbots miscount letters on unusual words. They read huge amounts of text, including lists of words, spelling bees, puzzles, and explanations, so they pick up associations such as which letters tend to form a word. This learned knowledge is patchy. They are good at common words and typical questions and less reliable at unusual ones, or at tasks requiring careful step-by-step counting.
There’s another factor. A language model produces answers by predicting plausible next text. The sentence “There are two r’s in strawberry” can feel plausible because people often focus on the “rr” in the ending “berry,” treating the double letter as one. The model may echo that human quirk.
Ways to get a correct answer
When AI chatbots miscount letters, you can often fix the problem with a better approach.
- Ask the model to spell the word out letter by letter first, then count. Separating the letters forces each one to become its own token.
- Request step-by-step reasoning. Models with more deliberate reasoning modes tend to do better on such tasks.
- Use a tool. If the system can run code, a one-line program counts letters perfectly every time.
- Double-check with a simple search or text editor when accuracy matters.
Newer models have improved on this specific puzzle, so AI chatbots miscount letters less often than they did, partly because the example became famous and partly because reasoning methods have advanced. But the underlying issue remains for other character-level tasks.
Related things that trip models up
The same root cause, behind why AI chatbots miscount letters, explains a family of oddities, and a related problem shows up when AI pictures mangle letters and words:
- Reversing a word or sentence character by character
- Counting syllables or the number of words in a long passage
- Writing in strict formats such as an acrostic or a fixed number of letters
- Arithmetic on long numbers, which tokenization can split awkwardly
- Rhyme and wordplay in less common languages
None of these show a lack of “intelligence” in the casual sense. They show a mismatch between how the system is built and how a task is framed.
What it says about trusting AI
When AI chatbots miscount letters, the failure is easy to spot. Harder failures are subtler, which is why it helps to understand why AI chatbots refuse and who decides too.
The strawberry story is a handy reminder to think of AI systems as strong at some things and surprisingly weak at others. The unevenness, such as when AI chatbots miscount letters, can be jarring because we expect competence to be smooth, as it is in humans. A person who writes well can spell. A model can write well and fail at spelling.
That’s why, when AI chatbots miscount letters, it is wise to match the task to the tool. Use language models for drafting, summarizing, brainstorming and explaining. Use calculators, spreadsheets and code for precise counting. Verify facts that matter, especially names, numbers and quotes.
A small experiment for you
Test why AI chatbots miscount letters yourself. Try three prompts with any chatbot: ask for the number of a letter in a long word, then ask it to spell the word with hyphens between letters and count again, and finally ask it to write a short script to do the count. Comparing results shows the effect of tokenization in a way no article can.
The takeaway
The reason AI chatbots miscount letters is a window into how language models read: in chunks, not letters. Once you know that, the failures become predictable and the fixes easy. Ask for step-by-step work, bring in code for exact tasks, and enjoy the fact that the world’s most hyped technology can still be tripped up by a fruit. Knowing the cause turns a funny failure into a practical lesson about picking the right tool for the job.
