Why AI Can't Count the R's in Strawberry
By Chatday Editorial Team ·
Ask a chatbot to plan your wedding and it nails it. Ask how many times the letter R shows up in “strawberry” and there’s a good chance it says two. The real answer is three. This is one of the strangest things about modern AI: the same tool that can explain quantum physics face-plants on a task a six-year-old can do with a pencil.
It isn’t a bug, and the AI isn’t dumb. It’s a side effect of how these systems read in the first place, and once you get it, a lot of AI’s weird little failures suddenly make sense.
The famous strawberry fail, explained
“How many R’s are in strawberry?” became the internet’s favorite way to embarrass AI. And for a while, almost every chatbot got it wrong, confidently answering “two” with the cheerful certainty of someone who has never been wrong in their life.
Spell it out yourself: s-t-r-a-w-b-e-r-r-y. There’s one R near the start and a double R near the end. Three total. Easy for a human, because we look at the actual letters. The AI doesn’t get to do that, and to see why, you have to look at how it reads.
AI reads in chunks, not letters
Here’s the part almost nobody knows. Before an AI model reads your message, the text gets chopped into pieces called tokens. A token is usually a common word or a fragment of one. The word “strawberry” often gets split into something like “st,” “raw,” and “berry.”
Now count the R’s the way the model does. The chunk “st” has none. “raw” has one. “berry” has two R’s, but the model sees it as a single lump, not as b-e-r-r-y. So the letters get buried inside chunks, and the model has to remember roughly how each chunk is spelled instead of reading it off the page.
One researcher put it perfectly: it’s like trying to count the threads in a rope without untwisting it first. The model sees the woven rope (the token), not the individual strands (the letters).
Why the same quirk breaks simple math
This is also why AI can breeze through a philosophy essay but trip over 4-digit multiplication. Numbers get chopped into chunks too, and not even consistent ones. The number 12345 might become “123” and “45,” so the model never really lines the digits up the way you would on paper.
On top of that, an AI isn’t calculating. It’s predicting the most likely next thing to say based on everything it has read. That works beautifully for language, where “close enough” is fine, and badly for arithmetic, where the answer is either exact or wrong. The model is built for probability, and math wants precision. Those two things don’t always get along. It’s a cousin of the reason AI can confidently make things up: a smooth, plausible answer is not the same as a correct one.
So why do some models get it right now?
Here’s the twist. Ask most chatbots about strawberry today and they’ll say three. The question got so famous that the correct answer is now baked into what they learned from. The models didn’t suddenly gain letter-vision. They memorized the punchline.
Swap the word and the mask slips. In one careful 2025 test, researchers asked how many B’s are in “blueberry.” One popular chatbot got it wrong roughly four times out of ten, confidently claiming three when there are two, complete with little check marks next to its work. Another model got it right every single time by patiently going letter by letter. Same task, wildly different results, which tells you this is less an iron law of AI and more a question of whether a given model bothers to slow down and check.
That’s the honest version of the story. The chunking problem is real, but it isn’t the whole thing. Even AI that shows its reasoning can miscount, because it’s pattern-matching its way through the steps rather than truly perceiving each letter. Newer “thinking” models are much better at this than the quick ones, which is a good example of fast AI versus smart AI in action. Better, though, is not the same as guaranteed.
What AI is great at, and what to double-check
You don’t need to memorize the plumbing. You just need a feel for where AI is rock-solid and where you should make it show its work.
| Task | How AI does | What to do |
|---|---|---|
| Writing, summarizing, brainstorming | Excellent | Trust it, then skim |
| Explaining a concept in plain words | Excellent | Trust it |
| Counting letters or characters | Shaky | Ask it to spell it out |
| Long or exact arithmetic | Shaky | Ask for step by step, or use a calculator |
| Precise details in a long document | Mixed | Ask it to quote the source line |
The pattern is simple. When the job is language, let it flow. When the job needs exactness, slow it down or switch to a model that’s careful.
How to actually get the right answer
Three moves fix almost every counting or math slip:
- Make it show its work. “Go letter by letter” or “solve this step by step” pushes the model to lay things out where it can see them, instead of blurting a guess.
- Ask twice, or ask two models. If accuracy matters, don’t trust the first snap answer. A second opinion from a different model catches a surprising number of these.
- Use the right model for the job. The fast, chatty models are built for speed. For anything fiddly and exact, a slower thinking model is worth the extra few seconds.
That last one is easier when you’re not locked into a single AI. You can line up a few of the top models on the same question and see who actually gets it right.
The takeaway
AI’s letter-counting fail isn’t a sign that it’s secretly stupid. It’s a window into how these tools actually work: reading in chunks, predicting instead of calculating, brilliant with language and clumsy with the literal. Once you know that, you stop being surprised and start steering. Ask it to slow down, make it show its work, and when the answer really has to be right, get a second model to weigh in.
The best way to build that instinct is to poke at it yourself. Ask a model to count the R’s, then ask it to spell the word out first and watch the answer change.