OpenAI. (2026). An AI analysing the word strawberry and tokenisation [AI-generated image]. ChatGPT.
How many r's are in the word strawberry?
We are living in a strange moment: AI can help solve problems that challenge some of the world's best mathematicians; it has reached gold-medal standard on International Mathematical Olympiad problems; and new benchmarks now test it on research-level maths. However, ask it a simple question – 'How many r's are in the word strawberry?' – and things can suddenly get tricky.
For humans, this is easy. We look at the word and count: one, two, three. But for many AI systems, this kind of task has been surprisingly difficult. Some models have answered two. Some have answered three, sometimes only after being corrected. Others can count the r's in strawberry correctly but fail when the word is intentionally misspelled. The 'How many r's are in strawberry?' question first became widely discussed across social media and tech forums around mid-2024, but it has remained a popular AI test into 2026, especially as newer models began getting the original example right while still sometimes failing on similar character-counting tasks.
This tiny puzzle reveals something much bigger; AI does not 'see' words, numbers and symbols in the same way we do.
Why does AI struggle with something so simple?
To understand this, we need to talk about tokens.
Large language models (LLMs) do not usually read text letter by letter. Instead, they break text into chunks called tokens. A token might be a whole word, part of a word, a punctuation mark, or even a space plus a word.
So, while we see strawberry as:
s t r a w b e r r y
an AI model may process something more like:
straw + berry
or another chunked version depending on the system.
This matters because if the model is not directly looking at each individual letter, then counting letters becomes less natural. It is a bit like asking someone to count the number of chocolates in a box while showing them only grouped sections of the box, not each chocolate separately. AI might still infer the answer, but it is not counting in the direct, one-by-one way we do when we look at the word letter by letter.
That does not mean artificial intelligence is not intelligent; it means its intelligence works differently.
The strange gap in AI ability
AI has now been credited with a breakthrough on an 80-year-old geometry problem posed by Paul Erdős but makes a mistake on a basic counting task. That feels contradictory.
LLMs are very good at predicting patterns. They learn from enormous amounts of text and become skilled at producing responses that are likely to fit your question or request. In many scenarios, this looks like 'reasoning', and some thinking models may even involve reasoning-like steps. But this is different from the kind of reasoning and thinking we would use. Particularly in the context of maths, one typo, one miscounted number, one missed symbol, or one wrong assumption can change everything.
So, can AI really do maths?
The answer depends on what we mean by 'do maths'.
If by 'doing maths' we mean recognising patterns, solving familiar types of problems, explaining methods in simpler language, or suggesting possible routes to a solution, then yes – AI can be very powerful. But 'possible' does not always mean 'correct'. If by 'doing maths' we mean understanding every step and guaranteeing that the answer is correct, then we need to be more careful. AI's answers still need checking, especially when precision and proof matter.
What the strawberry example teaches us
The strawberry example is interesting because it seems so simple, yet it can be surprisingly difficult for AI. It shows that AI can produce confident, convincing-looking 'reasoning' steps even when the answer is wrong. This is especially important for us to be aware of in maths education. If students use AI only to get answers, they may miss the point of learning. The accuracy of AI-generated answers to maths problems may continue to improve as systems and models develop, but if students focus only on the final answer rather than the reasoning behind it, they may not develop the deeper understanding that mathematics requires.
In our most recent research, we have observed that many students are already using AI in more productive ways. Rather than simply asking for an answer, they ask AI to help them understand a complicated concept in more accessible language or with concrete examples, compare different methods, or generate a similar question so they can test their understanding of the original problem. In these cases, AI can act as a valuable supporting tool, even without the immediate guidance or presence of a maths teacher.
I believe the key skill is not simply knowing how to craft the perfect prompt, but knowing how to question the help AI gives.
A better way to use AI for maths
Rather than asking, 'What is the answer?', or simply copying and pasting a maths question and hoping AI will solve it, we could use AI to explore the problem more thoughtfully:
'Can you explain this step by step?'
'Can you show me another method?'
'What assumptions are being made here? And why?'
'How can I check whether this answer is correct?'
'Can you give me a similar problem to practise and test my understanding?'
These questions turn AI from an answer machine into a learning partner.
Can AI really do maths, or is it just guessing cleverly?
Maybe the honest answer is: sometimes both.
The famous strawberry example is not proof that AI is bad at maths. Instead, it is a reminder that AI does not always handle text and maths-based tasks the way we do, and it reminds us why humans still need to stay in the loop: we notice the mistake, question the explanation, discuss what went wrong, and find better ways to improve and use the technology.
The future of maths education is not humans versus AI. It is humans learning how to work with AI wisely. AI may help us solve difficult maths problems, but maths education must still protect the most important skill of all: learning how to learn, reason and think. And sometimes, that begins with something as simple as counting the r's in strawberry.