Ask a modern AI chatbot to write a poem, summarize a contract, or explain a historical event, and it often does a genuinely impressive job. Ask the same chatbot to multiply two five-digit numbers, and it can still get the answer wrong with total confidence. That gap surprises people, because we are used to computers being bad at language and good at arithmetic. With chatbots it is the other way around, and the reason says a lot about what these systems actually are.
A Language Model Is Not Running a Calculator
A large language model is trained to continue text. Given everything written so far, it predicts what is most likely to come next, then repeats that one piece at a time. When it answers a maths question, it is producing that answer the same way it produces the next word of a sentence: by pattern, not by procedure.
For small sums this works fine, because expressions like two plus two appear constantly in training data and the pattern is overwhelming. For arbitrary large numbers there is no memorised pattern to lean on. The model has never seen your specific multiplication before, and it has no internal calculator to fall back on.
Numbers Get Chopped Up Before the Model Ever Sees Them
There is a second problem that happens before any reasoning starts. Text is split into tokens, and number tokens rarely line up with place value. One number might be split after three digits, a slightly longer one after four. The model does not reliably receive a clean ones column, tens column and hundreds column the way you see them when you write a sum out on paper.
A person doing long multiplication leans heavily on that alignment. The model has to infer it from chunks that were carved up for the convenience of text compression, not arithmetic. It is a bit like being asked to add two numbers that someone has read aloud to you in uneven groups.
Long Problems Need Steps, and Small Errors Compound
Multiplying two five-digit numbers is not one operation. It is a stack of partial products and carries that have to be tracked and combined in the right order. A model does a fixed amount of internal computation per token it produces, so it cannot silently loop through twenty steps before answering. If it tries to give the result in one shot, it is compressing a long procedure into a single prediction.
That is why accuracy falls off a cliff as numbers get longer. Get one carry wrong and the whole answer is wrong. Even a small per-step error rate becomes a likely failure once there are enough steps.
Why Showing the Work Helped So Much
The first real improvement came from getting models to write out intermediate steps instead of jumping to an answer. Working through a problem in the open does two things: it gives the model more computation to spend, since it produces more tokens, and it creates a scratchpad it can read back as it goes rather than holding everything in one pass.
This is why so many assistants now display their reasoning before a final number. It is not decoration. The visible steps are part of how the answer gets computed.
The Bigger Fix: Hand the Sum to an Actual Calculator
The change that mattered most was letting models stop doing arithmetic themselves. Rather than predicting a result, the assistant writes a small piece of code or calls a calculator tool, runs it, and reports what came back. A guess becomes a deterministic computation.
Most current assistants do this for anything beyond trivial sums. It is worth understanding what that means: AI maths got dramatically more reliable not because the language model learned mental arithmetic, but because it learned to recognise a maths problem and delegate it.
The Part That Should Still Make You Careful
A chatbot has no calibrated sense of its own uncertainty about a number. It states a wrong figure in exactly the same steady tone as a right one, with no hedging and no error bars. In a piece of writing, a small mistake is a style issue you can edit. In a number, it is simply wrong, and nothing in the output flags it.
Tool use narrows this but does not eliminate it. A tool call can fail quietly, the model can feed it the wrong inputs after misreading a paragraph, or it can compute the right value and then mistype it in the summary sentence. The arithmetic is solid; the handoffs around it are still language.
How To Get Numbers You Can Actually Use
A few habits make a real difference. Ask the assistant to show its steps or to use a calculator or code rather than answering directly. Hand it clean figures instead of making it pull them out of prose, because misreading the input is a common failure. Sanity-check the scale of the result yourself, since an order-of-magnitude error is much easier to spot than a wrong final digit.
And for anything that touches money, health or a legal obligation, move the calculation into a tool you control, such as a spreadsheet, and use the chatbot to explain the method rather than to produce the number.
Bottom Line
Fluency and accuracy are separate capabilities, and a chatbot has far more of the first than the second where numbers are concerned. Arithmetic was never native to these systems; it was bolted on, and it works well now mostly because the model has learned when to stop guessing and reach for a calculator. Until you have checked it, treat any number a chatbot gives you as a draft rather than a result.
Related video
For readers who prefer to see the idea explained visually, this related YouTube video is included as a helpful companion to the article.
Image credit: https://images.unsplash.com/photo-1773332585815-f106a5d6ed6c. Used as a copyright-free/royalty-free editorial image with source attribution.
Video credit: Related YouTube explainer on ChatGPT and algebra via YouTube. Embedded for reader education and context.
Editorial note: This guide is written for everyday readers. It focuses on practical understanding, safe choices, and clear trade-offs rather than hype.