Can AI actually do maths?
Models handle digits as chunks of text, which is why arithmetic drifts. Four steps to a maths answer you can trust, plus one check.
The VAT on an invoice comes back in about a second, set out neatly, with the stages summarised underneath. The calculation looks convincing. Sometimes it is right. Sometimes the tax was worked out on the full amount instead of the discounted one, and every number below has inherited that while looking just as tidy.
The honest answer to the question in the title is a split one. On the method, setting the problem up, naming the right formula, turning a word problem into arithmetic, saying which line of your own working went wrong, these models earn their place. On the arithmetic itself, the grinding through of digits, they are less dependable than the calculator that came free on your phone. Once you know why the split is there, the four steps below stop looking fussy.
Why the digits are the weak part
Underneath, the model is doing one thing over and over: it reads everything so far and adds the next piece of text. That is the whole mechanism, and it is also the source of the confident wrong answer.
Those pieces are chunks of characters rather than digits. For ordinary words they fall roughly where syllables fall; for numbers they fall wherever the model's vocabulary happens to split them, so a figure like 1,375 may arrive as two or three chunks, comma included. Set out a long multiplication on paper and you are holding a column of digits, each in a place that means something. The model is handling chunks of text that contain numerals, with no columns and no place value underneath them.
Then there is length. Long multiplication runs as a chain of small decisions: multiply, carry, shift, add, repeat. Every link is another chance to drift, and an early drift is invisible by the end, because the model carries on from the number it produced rather than the correct one. The wrong total arrives in the same unbothered register as a right one.
One more thing shapes the steps below. No calculator runs behind the answer: we do not execute code on your behalf, and the model is not dialling out to a maths engine. When a model shows its working, it is writing a description of a calculation, and a description can carry a false line while reading immaculately.
Treat any number that will cost you something, on an invoice, a tax return, a dosage, a loan repayment, as a draft until you have checked a step yourself. The four steps make that check cheap. None removes the need for it.
Step one: hand over the whole problem
Plenty of maths answers go wrong before any arithmetic happens, because the problem arrived incomplete. Write the numbers out in full rather than rounding on the way in, name the units or the currency, and say which operation comes first if the order is arguable. "Apply the discount, then the VAT on the discounted amount" removes a guess the model would otherwise make in silence, and might make differently next time.
If the problem is on paper, a photograph saves the typing. 46 of our 59 models accept an image, and the phone's attachment menu has a camera option, so a snap of a worksheet, a bill or a page of your own working goes straight into the message. Free accounts get one attachment per message, up to 10MB, three a day.
Before you trust anything built on a photograph, ask the model to read the figures back as a list. A 7 misread as a 1 gives you an answer that is internally perfect and completely wrong, which is the hardest kind to catch later.
Step two: turn the thinking up so you can read it
You can set the depth of thinking yourself on 41 of our 59 models. On a maths question the change is concrete: the model writes a working pass before the total, and you can read it. It appears above the reply as a "Thinking..." line while it runs, then collapses into a block labelled "Reasoning", or "Reasoning · 3 steps" if it took more than one. That block is what you want in front of you here, because a total on its own gives you nothing to check.
Where the control sits depends on the screen. The phone keeps a pill in the row under the message box on every model that reasons. The website starts you without that button at all, because Intent Detection ships on for every account, and while it is on the app reads each message and settles search, image generation and, absent a level from you, the depth of thinking. Its switch is in Settings, then Tool Preferences. Turn it off and the button appears. On a free account the control shows where it stands and will not move; tapping it opens the upgrade prompt instead of the menu.
Step three: ask for the working, then check one line
Ask for the stages rather than the answer: "show each step with the intermediate figure, and say which operation you applied first."
Take an invoice of 2,450 with a 12 per cent early payment discount and 21 per cent VAT on the discounted amount. Asked cold, with no working pass, the usual reply is one confident total and a sentence of justification. Sometimes the 21 per cent has gone on the full 2,450 instead of the 2,156 left after the discount, and the gap that opens is the VAT on the 294 the discount took off, not a rounding difference. Sometimes an intermediate value gets rounded to something tidy, 2,156 becoming 2,150, and the rounded figure carries into everything after it. With the working pass on, the typical reply sets the stages out separately, states the order it chose, and carries the unrounded intermediate through. That is what makes the next move possible.
Now pick one line and do it yourself, usually the first, since everything below depends on it. One sum on a calculator, and it either matches or it does not.
Working that exists is not working that is right. A model can write steps that each read plausibly with an arithmetic slip inside one of them, or produce a correct sequence and then state a total that does not follow from it. Read what the lines say rather than being reassured that there are lines.
Step four: put the same question to a second model
This check is worth more on maths than on almost anything else, because the answer is either right or it is not. Two models agreeing on a paragraph of advice tells you little, since both can be bland in the same direction. Two models reaching the same total is real evidence. Two different totals mean either one of them has slipped or they read your question in two different ways, and you found out for the price of one message.
Two routes. Switch model inside the conversation and ask again, which keeps the thread and everything in it, so there is nothing to paste. Or branch, which opens a new thread from that point and leaves the original where it was. There is no side-by-side view in the app, so you read them one after the other.
Pick your second model from a different provider, not another member of the same family. Different families were built differently and drift in different places, which is the point of asking twice.
Where a chatbot beats a calculator
Plain arithmetic belongs on a calculator, and nothing above argues otherwise. What a chat window adds is everything around it: the formula you had forgotten, the word problem you could not translate, the explanation of why the answer is what it is, and the pass over your own working that finds where you went wrong.
Checking an existing calculation is useful too. Photograph your working, ask which step is wrong, and read the reply with step three's warning in mind. A calculator will never tell you that line four is where it fell apart. Finding that out is a reading job, and reading is what these models were built for.