
A short question is worrying a lot of people who love numbers: what happens to maths when AI gives the answers and cannot say why they work? It is a big question, and it touches anyone who studies the subject, from GCSE students to professors.
The question came up again this month when the mathematician Steven Strogatz was interviewed about his new book, Big Math, which he wrote with Alex Townsend. In this guide we explain the main ideas in our own words, add the context a student needs, and finish with practical advice. It is a news explainer, not a copy of the interview, so if you want to hear it from the author, you can read the original at Ars Technica.
What is "big maths"?
Big maths is not about big numbers. It is about problems that are too large for a human brain to handle, even though the ideas inside them can be simple. Think of a system with billions of unknowns, like the one that helps plan a flight path, scan a patient or forecast the weather.
The idea is older than computers. Long ago, people in China used a counting board with rows and columns of bamboo sticks to solve several equations at once. That board was a hand-held version of a matrix. So handing maths over to a tool is not new. What changes is the size of the job.
If you have solved a pair of linear equations at school, you have already done a tiny version of this. Computers do the same thing with millions of equations, using methods such as Gaussian elimination. A fun surprise from the interview is that infinity is not the hardest thing here. Calculus handles infinity neatly. A messy problem with a billion moving parts is much harder.
Servant, scout, oracle
A helpful way to see the story is in three stages.
- Servant. The computer does exactly what we tell it, only faster. It is like a calculator, a telescope or a chisel. We understand every step.
- Scout. From the 1950s, people sent computers into places where nobody knew the answer, then asked them to report back. The results surprised us, but we could still work out why afterwards and build a theory.
- Oracle. This is where we are heading. The AI says what is true or what will work, and we cannot always follow its reasoning. It seems to be right, but it cannot teach us the pattern it used.
Two everyday examples were given. An online translator can turn one language into another very well, yet it has not given us a new rule of grammar. And the AI that predicts the shape of proteins has helped scientists with far more proteins than before, without teaching us new chemistry.
What the first scout found
The "scout" stage began with a famous early experiment run by Enrico Fermi and colleagues. The calculations were done by Mary Tsingou. They wanted to see how energy spreads out among vibrating atoms, which is the same as asking why things drift towards disorder.
Pencil and paper could not follow the problem far enough, so they used a computer. Everyone expected the energy to spread out and stay spread out. Instead, after a while, nearly all of it (about 98 per cent) came back to where it started. Nobody saw that coming. It led to a whole new area of maths, called soliton theory, and it showed that human intuition can fail once a problem is not a simple straight-line pattern. The computer went where we could not walk, and then people made sense of what it saw.
That is the happy version of the story: the machine finds, and the humans understand. The worry is that this second step may not always be possible.
How do we trust an oracle?
If a system cannot show its reasoning, why believe it? There are three answers people give.
- Check it by testing. Medicine works this way. Doctors do not need to know exactly how a drug works before using it. They run careful trials and count how often it helps. An AI that is right nearly every time can be tested in the same way.
- Hope it becomes a good teacher. An AI that is hard to follow today might explain itself better in ten years.
- Accept that some things may stay out of reach. A few systems may be too big for any person to follow, even with help.
The interview also mentioned an idea from the scientist Stephen Wolfram, who suggests that most of nature may have no short, tidy explanation at all. The only way to know what happens is to run it and watch. Newton's laws of motion, with a few simple rules covering the planets, may be the lucky exception. Nobody knows if that is right for nature, but it is true for many computer simulations.
What might we lose?
This is where the conversation turns more serious, and it is worth being honest about it.
Certainty. Maths is special because a proof, once checked, stays true forever. Euclid's proofs from about 300 BC are still correct today, while almost nothing from the science of that time is. If maths starts to rely on "it is right nearly every time", it would feel more like science or medicine, and some pure mathematicians would find that a real sacrifice.
The struggle. Anyone who has stared at a problem for an hour and then suddenly seen it knows the joy of the "aha!" moment. Along the way you try simpler versions, you invent new tricks and you learn far more than the final answer. If we only collect answers, we skip the part that makes us better at maths.
The human voice. Mathematicians have styles. You can tell them by the questions they ask and the way they explain. Writers and artists feel the same worry about work with no personality behind it.
There is also a very human cost. Young students who dream of a career in maths are hearing that the subject is "over". The interview does not pretend this away, and neither should we.
Reasons for hope
The story is not all gloomy, and the author says so himself.
- AI can be a great explainer. Used well, it can help more people understand more maths than ever before.
- Small discoveries still count. You do not need to win a top prize to enjoy finding something out for yourself. People do puzzles, Sudoku and chess for the joy of thinking, and maths can be the same.
- AI is weak at asking questions. Good questions are the hardest thing to teach new researchers, and so far AI is not good at it. Learning to ask a sharp question is a skill that stays valuable.
- The benefits are real. Better medicines, cleaner energy and new physics are all possible with these tools.
The book ends with the old story of Faust, who makes a deal for greater power. In the original tale by Goethe, the ending is less bleak than people expect: if we keep striving for something bigger than ourselves, there may be a way to be redeemed. It is still described as a risky game to play.
What this means for students
None of this changes how your exam is marked. You still need to show clear working, and you still lose marks for a lucky guess. Here is a simple plan for using AI and solvers well.
- Learn the method first. Work through a guide such as how to solve quadratic equations or how to differentiate on paper.
- Then use a solver to check. Type the same question into the step-by-step solver and compare each line with your own.
- Find the first different line. This is where you learn the most. Fix that step, not the whole question.
- Ask "why?" out loud. If you cannot explain why a step works, practise it with practice questions until you can.
- Look at the bigger picture. If you are curious about how real computers solve huge problems, browse numerical methods, linear algebra and differentiation. They are the A-level and university roots of "big maths".
The tools are getting more powerful, so understanding matters more, not less. If you know why an answer is right, you will always be able to tell when a machine gets it wrong.
Frequently asked questions
What does it mean when AI is an "oracle" in maths?
An oracle is a tool that gives you answers that are usually right, but cannot explain how it got them. In maths this would mean a system that finds results we can check but may not fully understand.
Will AI replace mathematicians?
Nobody knows yet. AI is already helping to find new results, but people still choose which questions matter, judge which ideas are interesting and explain them to others. Many mathematicians expect the job to change rather than disappear.
Should I still learn methods if AI can solve the question?
Yes. Exams ask you to show working, and you cannot spot a wrong AI answer unless you understand the method. Use a solver to check your steps, not to skip them.
Can I trust an answer I do not understand?
Treat it like a claim to be tested. Check it with a different method, substitute it back into the question or compare it with a trusted solver that shows every step.
Is maths still certain if computers help to prove things?
A proof that people can read and check keeps its certainty. The worry is about results that are too long or too tangled for any person to follow. Those would need new ways of being checked.