It became a famous party trick at AI's expense: ask a powerful chatbot how many times the letter "r" appears in "strawberry," and watch it confidently answer "two." Ask it to multiply two biggish numbers and it may hand you something close but wrong. For a machine, getting these school-level tasks wrong feels absurd — until you understand what the model is actually looking at. The same quirk explains both, and it says something important about what these systems are and aren't.

The model can't see letters

Here's the surprise: a language model doesn't read text as a string of letters. Before anything reaches the model, your text is chopped into tokens — chunks that are often whole words or fragments of words. "Strawberry" might arrive as just two or three tokens, not the ten individual letters s-t-r-a-w-b-e-r-r-y. The model never sees the letters at all; it sees an opaque chunk that means strawberry.

So when you ask it to count the r's, you're asking it about something it can't directly look at — like asking someone to count the letters in a word you've only ever said aloud, never spelled. It can sometimes get there by recalling, from training, how the word is spelled — but it's reconstructing the spelling from memory, not reading it off the page. Get an unusual word, and the trick falls apart. The model isn't stupid; it's been handed the wrong representation for the job.

Maths is the same problem in a different coat

Arithmetic exposes a related gap. A language model has no calculator inside it. It doesn't compute 7,823 × 461; it predicts the next token, the same way it predicts the next word in a sentence. For small, common sums it has effectively memorised the answers from training text, so it's reliable. For large or unusual ones, it's pattern-matching its way to a plausible-looking number — and "plausible-looking" is not the same as "correct." Numbers also get tokenized in awkward ways, so the digits it's reasoning over may not even line up the way you'd expect.

The deeper point under both examples is the same: a language model is a brilliant pattern-predictor over text, not a symbolic engine that manipulates characters and digits exactly. Spelling and arithmetic are precisely the kind of exact, symbol-level operations that this design is worst at — which is why a system that can write you a working program or a passable sonnet can still trip over a five-year-old's spelling test.

How modern AI gets around it

The fix in practice isn't to make the model "try harder" — it's to give it the right tool. Ask a capable AI assistant to count those r's or do that multiplication using code, and it will often write a tiny program, run it, and read back the exact answer. This is tool use: the model recognises a task it's bad at and hands it to something that's good at it — a calculator, a Python snippet, a search. A second trick that helps is making it slow down and spell the word out letter by letter, or work the sum digit by digit; forcing the steps into the open often gets it to the right place where a snap answer wouldn't.

The takeaway

The lesson isn't "AI is dumb at maths." It's more useful than that: these models have a specific, predictable blind spot for exact, character- and digit-level operations, because they don't perceive text the way we assume they do. So treat raw mental arithmetic and letter-counting from a chatbot with suspicion — and when the answer has to be exact, make sure it's using a tool to calculate, not its intuition to guess. Knowing where the blind spot is turns a baffling failure into a manageable one.