Why Is AI Bad at Math? ChatGPT Mistakes Explained

Quick answer: AI chatbots get math wrong when they misread the question, mishandle numbers, or make an invalid reasoning step, and their answers are not automatically checked. Accuracy depends on the model, the task, how much reasoning it does and whether it uses tools. A calculator or code tool reduces calculation errors, but cannot fix a problem set up wrongly. And no, 9.11 is not bigger than 9.9: as decimals, 9.9 = 9.90, which is greater than 9.11.

Sources checked November 22, 2026.

In July 2024, users shared a chatbot answer claiming that 9.11 is bigger than 9.9. Two years later, AI systems were reporting perfect olympiad scores and contributing to research proofs. Both happened, but with different systems under different conditions. Once you understand where AI math errors come from, they stop being mysterious, and you can catch them quickly.

Is 9.11 Bigger Than 9.9?

No. Write both with the same number of decimal places: \(9.9 = 9.90\). Both numbers have 9 ones; 9.90 has 9 tenths and 9.11 has only 1 tenth, so \(9.9 > 9.11\) (see comparing decimals). One possible reason a chatbot slipped is that in other settings, such as software versions or book chapters, 9.11 does come after 9.9. A screenshot alone cannot tell us the cause, but the example shows how a fluent answer can still be wrong.

5 Reasons AI Makes Math Mistakes

ReasonWhat happensTypical mistake
1. Number representationNumbers are encoded as tokensA wrong digit or carry in a long calculation
2. Learned arithmetic can failProcedures learned from data are not guaranteedAn answer that looks plausible but is off
3. Sensitivity to wordingAn extra or changed detail shifts the patternUsing an irrelevant number
4. Wrong setupThe question is translated into the wrong equationCorrect algebra on the wrong problem
5. Confident wordingWrong steps sound as sure as right onesA smooth explanation of a false result

1. How Models Represent Numbers

Language models read text as tokens. Depending on the tokenizer, a number like 48,317 may become individual digits or groups of digits. When you multiply on paper, you line up place values and carry from column to column. A model has to handle place value through its learned representation, which usually works but can fail when a calculation needs many consistent carries. Tokenization is one contributing factor, not the whole explanation.

2. Learned Arithmetic Can Fail

A language model is not a calculator, but it is not just guessing either. Anthropic’s 2025 interpretability research found that Claude 3.5 Haiku adds numbers using internal pathways, one estimating the rough size of the answer and another computing the last digit, and that its own explanation of how it added did not match what happened inside. Learned procedures like this can be impressive and still break on long or unusual inputs.

Older models showed this clearly. In the 2023 “Faith and Fate” study, Nouha Dziri and colleagues found that GPT-4’s accuracy on multi-digit multiplication dropped sharply as the numbers got longer. That was a historical experiment; current reasoning models do much better by writing out partial products. For reference, the correct value of one such product is \[ 4{,}817 \times 3{,}906 = 18{,}815{,}202. \]

3. Small Changes in Wording Can Break It

In October 2024, Apple researchers led by Iman Mirzadeh published GSM-Symbolic, built from grade-school word problems. Changing the numbers in the problems lowered accuracy for many models, and changing only the names usually had smaller effects. In a variant called GSM-NoOp, they added one irrelevant sentence, for example that five of the kiwis picked were “a bit smaller than average.” Some models subtracted those five, and the worst-affected models scored up to about 65 percentage points lower than on the standard set. Newer reasoning models are more robust, but distracting details are still worth watching for.

4. Setting Up the Wrong Problem

A model can do every step correctly and still answer a different question, for example by reading “at least” as “exactly” or by missing the word “each.” A calculator does not help here: if the equation is wrong, the exact answer to it is wrong too.

5. Confident Wording Is Not Proof

A chatbot writes a wrong step in the same smooth tone as a right one. If you push back, older models often caved in to a wrong correction. A polished explanation is not evidence; checking is.

What Reasoning Models and Tools Change

SetupWhat it addsWhat you should still check
Model answers directlyFast learned answersEvery step
Reasoning modelWrites out intermediate work and checks itselfThe setup and key steps
Model runs code or a calculatorA separate exact computationThat the right expression and inputs went in
Formal proof checker (Lean)Machine-checked logicThat the formal statement says what was meant

How much do these changes help? In our own small test on November 22, 2026, ChatGPT in Pro mode answered all 20 questions we gave it correctly, including long multiplication without tools, the strawberry letter count and a GSM-NoOp-style kiwi problem, and it kept its answer when we pushed back with a wrong one. That is a demonstration, not a benchmark; see the full ChatGPT math test.

Why AI Can Solve Hard Problems and Still Slip

Strong research systems are not a single chatbot reply. They explore several approaches, run calculations, revise their work and sometimes check proofs in Lean, as in Google DeepMind’s 2026 Aletheia research agent. A quick unaided answer in a chat window has none of that support. So the same family of models can contribute to research and still make a basic slip in a casual answer. For the wider picture, see Can AI Do Math? and how mathematicians use AI.

Famous AI Math Mistakes, and the Correct Answers

QuestionCommon wrong answerCorrect answerIndependent check
Which is bigger, 9.11 or 9.9?9.119.9Compare 9.90 with 9.11
A bat and a ball cost 1.10 dollars in total. The bat costs 1 dollar more than the ball. What does the ball cost?10 cents5 cents\(0.05 + 1.05 = 1.10\)
\(48 \times 52\)24,9602,496\((50-2)(50+2) = 2{,}500 – 4\)
How many letters r in “strawberry”?23s-t-r-a-w-b-e-r-r-y

These were widely shared failures of 2023–2024 chatbots; current reasoning models usually get them right.

5 Checks That Catch AI Math Errors

  1. Estimate first. \(48 \times 52\) is about \(50 \times 50 = 2{,}500\), so 24,960 is off by a factor of 10. See estimation.
  2. Substitute back. If AI says \(x = 3\) solves \(2x + 5 = 11\), check: \(2(3) + 5 = 11\).
  3. Reread the question. Did it answer what was asked, with the right units and conditions?
  4. Ask for a computation you can inspect. If it runs code or a calculator, check the expression it used.
  5. Ask “why?” at each step. A wrong answer usually hides in one line.

Common Misconceptions

“If AI can’t multiply, it can’t do math.” No. Calculation and reasoning are different skills, and modern systems are strong at both under good conditions.

“A calculator tool fixes everything.” No. It fixes arithmetic, not a wrongly set-up problem.

“Newer models never make these mistakes.” They make far fewer, but slips still happen, so checking remains a good habit.

Watch: Why AI Makes Math Mistakes

If you learn better by watching, these videos are worth your time.

Why Is ChatGPT Bad At Math? (SciShow). From June 2023: a clear explanation of how language models handle numbers. Its examples are historical.
AI can be bad at math, and stubborn (MindYourDecisions). Real examples of chatbots defending wrong answers.
ChatGPT is destroying my math exams (Dr. Trefor Bazett). A math professor tests a chatbot on his own exam questions.

Check Your Understanding

Five quick questions to see how much of this has stuck. Pick an answer and press Submit; you will see a short explanation either way.

Answer each question and press Submit to see your result right away.

1. Which is bigger, 9.11 or 9.9?
Found a bug? Let us know!
2. What are tokens?
Found a bug? Let us know!
3. What is 48 × 52?
Found a bug? Let us know!
4. In the bat-and-ball puzzle, what does the ball cost?
Found a bug? Let us know!
5. An AI uses a calculator but types the wrong equation. What happens?
Found a bug? Let us know!

Frequently Asked Questions

Why is AI bad at simple math?

AI chatbots can misread a question, mishandle numbers encoded as tokens, or make an invalid step, and their answers are not automatically checked. Reasoning modes and calculator or code tools reduce errors, but you should still verify the answer.

Is 9.11 bigger than 9.9?

No. As decimals, 9.9 equals 9.90, and 9.90 is greater than 9.11 because 9 tenths is more than 1 tenth. Chatbots that got this wrong may have mixed it up with version or chapter numbering.

How can AI solve olympiad problems but miss simple math?

Research and olympiad systems explore several approaches, run code and check their work, while a quick chatbot reply has none of that support. Different systems and settings give very different reliability.

How do I check if AI math is right?

Estimate the answer first, substitute the result back into the problem, check that it answered the actual question, ask it to use a calculator or code, and ask why each step is true.

Are newer AI models better at math?

Yes. Reasoning models that write out steps and use tools make far fewer errors, but they can still slip, especially on long calculations without tools or on oddly worded problems.

Keep Exploring

Sources and Further Reading

Related to This Article

What people say about "Why Is AI Bad at Math? ChatGPT Mistakes Explained - Effortless Math"?

No one replied yet.

Leave a Reply