Why does AI fail at counting letters?
AI doesn't see words the way you do.
For a long time, if you asked a chatbot "How many r's are in strawberry?", it would confidently say two.
There are three.
This is a model that can write a sonnet or debug code. So why couldn't it count to three?
Before the model reads anything, the text is chopped into chunks called tokens (we met them in "Why long chats make AI worse"). The model sees the chunks, not the letters inside them.
Tokeniser: the tool that does the chopping. Each company has its own, so the same sentence splits differently in different models.
Common words are often a single token. Rare words get split into pieces.
You see:
s t r a w b e r r y
The model might see something like:
str · aw · berry
Three blobs. Each blob is just a number to the model. Asking it to count the r's is like asking you how many brushstrokes are in a word you've only ever heard spoken.
Made-up split for illustration. Real tokenisers vary.
Numbers can split in odd places. "2025" might be one token while "9.11" becomes "9", "." and "11".
That's part of why models used to say 9.11 is bigger than 9.9. They saw an "11" and a "9", and 11 is bigger than 9.
It's also why you shouldn't trust it to do maths in its head. (More in "Don't let AI do maths in its head".)
Example splits for illustration.
Tokenisers are built mostly from English text, so English packs efficiently. Many other languages get chopped into many more pieces for the same meaning.
That matters because:
Newer models usually get it right. Partly they've been trained on the strawberry question itself, now that it's famous. Partly some models now "think" first and spell the word out letter by letter before answering.
But the underlying blind spot is still there. Try a less famous word.
For letter-level tasks, make it spell first. "Spell the word one letter per line, then count."
For counting or maths, ask it to use code. Code works on the actual letters and digits.
Remember limits are in tokens, not messages. A message full of pasted data can count as much as a whole afternoon of chat.