Why it gets sums wrong

You typed a sum into the AI and the answer came back in one second. Neat working, a clear final number, no hesitation anywhere. Then you checked it on your phone calculator and the number was wrong.

Nothing is broken. The same app that explains a whole chapter well can add two long numbers wrong. This lesson shows you why that happens, and how to stop a wrong number reaching your exam paper or your money.

It never saw your digits the way you do

When you add two big numbers, you write one under the other. Units under units, tens under tens. Then you carry. The model never does any of that.

Your message is cut into small pieces before the model reads it. Each piece is called a token. You met this idea in it is guessing the next word.

A long number gets cut too. Google researchers checked this in February 2026. Some models chop a run of digits into groups of three, counting from the left. Other models read the digits one at a time.

So 4,73,829 is not six digits sitting in columns. To the model it is a few chunks of text. The units place is not where you think it is.

That study tested plain addition of two numbers. Accuracy fell as the numbers got longer. Even at 20 digits, most of the models tested made a real number of mistakes.

The kind of mistake is the interesting part. Most of the errors came from two habits you already know. First, the digits of one number sat a few places out, exactly like a sum written with the columns not lined up. Second, carrying went wrong when a column added up to 9 or to 10.

Those are class 4 mistakes. The AI makes them with complete confidence.

Here is one more clue that the chunking is real. For one model, addition was clearly more accurate when the number of digits divided by three exactly.

So it is remembering, not working it out

A calculator has a rule for adding, and it follows that rule digit by digit. The model has no such rule. It writes the answer the way it writes a sentence, one piece at a time, choosing what usually comes next.

So it is doing something closer to remembering what a sum like this normally looks like. Small sums appear millions of times in what it read. 7 times 8 is easy for it. A 12-digit multiplication is not, because that exact sum has probably never been written down anywhere.

That is why its skill has such an odd shape. In 2025, and again in early 2026, AI systems reached gold medal level at the world school maths olympiad. The same kind of system can still add two long numbers wrong.

Being strong at hard maths and being safe at plain sums are two different things here.

Two minutes to prove this on your own phone

Do not take my word for it. Test it tonight, on whatever app you already have.

Two more cheap tests. Ask it to count how many times one letter appears in a word, then count them yourself. Researchers tried this with the letter D in the word DEEPSEEK. They ran that test on several apps. The apps did not agree with each other. Some answers were as high as six and seven for one word.

Then ask it which of two similar decimal numbers is bigger. Pick two that start with the same digit. Check its answer with your own eyes.

Run these before you trust it with anything that matters. One test you did yourself teaches more than ten pages of warning.

Two ways to stop it guessing at the sum

Two very different things can happen inside the same chat box.

In the first, the model writes the number straight out of its own head. That is the guessing described above.

In the second, the app writes a small program, runs it on a real computer, and reads the result back. That is a real calculator, and it is accurate. In the February 2026 study, one model was allowed to run code. It stayed close to perfect, because it worked most sums out with a short program.

Ask for it in plain words:

Use your code tool for every calculation.
Show me the code and its output.
Do not work the numbers out in your head.

Then look at the reply. If a small block of code appears, or a line saying it ran something, a real calculator did the work. If nothing like that shows up, the number came out of memory. Not every app and not every free plan has this tool, so how it searches, sees and uses tools is worth reading first.

One risk is left even then. The same study found the leftover errors came from two places. Either the model did not use the tool at all, or it used the tool and then copied the number into its answer wrongly. So read the last line of the reply against the tool output.

The second fix is thinking mode. This is where turning it on pays off most for the extra wait. It writes rough work before the final answer, and on sums and puzzles that helps. As of August 2026 it clears up many of the letter counting and long number failures. Turn it off again for simple questions, as what thinking mode really does explains.

If your app has neither, at least force the working into the open:

Solve this step by step.
Show every line of the working.
Put the final answer on its own line at the end.

The two ways a maths answer goes wrong

Two failures matter here, and they hurt you in opposite ways.

The first is a right final answer with wrong working. One line in the middle does not follow from the line above it. The model slipped, then quietly wrote the correct next number as if the slip never happened. Researchers have studied this and given it a name.

The danger is not the answer. It is that you copy that method into your notebook and carry it into the exam hall.

The second is neat working with a wrong final number. Every line reads well. The last number is wrong anyway. In a board paper or an SSC paper, that is your mark gone.

Remember what the working really is. It is a tidy explanation the model wrote for you. It is not a recording of how it reached the answer. Anthropic tested this in 2025. They gave a model a hidden hint, and the hint changed its answer. The model mentioned that hint in its written reasoning only about a quarter of the time.

So the rule is simple. If you are copying a method into your notebook, work out every single line yourself with a pen. Never assume line 4 follows from line 3 just because the last number matched the answer key. Build the habit in the two-chat test and three other checks.

Type the numbers so it can read them

Small changes in how you type a sum make a real difference.

Write big numbers with separators. Type 4,73,829 rather than 473829. Because the model chops digits from the left, the commas push it to break the number where you break it. Researchers measured this in 2024, and the answers got better.

Next, take the story out. A model that gets 36 times 6 right can slip when the same sum is wrapped in a story about a boy and his apples. Researchers found exactly that pattern. So keep the story for yourself and hand over clean numbers.

A good order of work is this. Ask the AI which sum you have to do. Then do the sum yourself.

Do not give me the final number.
Tell me in one line which calculation I have to do.
Use rupees and keep the wording simple.

Now the trap that catches people. Extra numbers that are not needed can break the answer. Look at this one.

“Ram bought 12 mangoes at Rs 15 each. Five of them were a bit smaller. How much did he pay?”

The five smaller mangoes change nothing. He paid Rs 180. Back in 2024, one extra sentence like that could pull a model’s score down badly. Most of that weakness has been fixed since. A small, cheap model tested in May 2026 scored almost the same with and without the extra sentence.

Mostly fixed is not always fixed, and a wrong answer looks exactly like a right one. The defence costs you one line:

Before you solve this, write the problem again in simple words.
List what is given, what is asked, and which number is not needed.
Then stop and wait for me to agree.

Read that list carefully. If it has quietly picked up the number that does not matter, you have caught the mistake before it cost you anything.

The rule for money, and what comes next

None of this makes AI useless for maths. Used properly it is a patient teacher, which is the whole point of making AI your maths teacher. It just cannot be your teacher and your checker at the same time.

Checking is your job, and it is a skill you can learn. The next lesson gives you seven ways to check a maths answer with no tool at all. Those checks work in the exam hall too, where you will have only a pen.