Scientifically Speaking: As AI cracks hard Math, how will my son's generation thrive?

My son is a ninth grader who has recently become interested in using computers to answer biological questions. Computational biology is young as scientific disciplines go, but it is now central to modern biology and medicine, and biologists often test ideas on a computer before taking them to the lab or the field. He enjoys it and does not worry about AI the way my generation does (job losses, loss of cognitive abilities, existential threats; well, you get the picture). More specifically, I worry about how much of the work will still be done by people when he reaches college. AI can already write code, sort datasets and assemble the steps of an analysis, and some automated laboratories combine AI with robots to run selected experiments.

For a preview of where this may lead, look at mathematics, where people have started talking about a "mathpocalypse". The moment that shook the field came on September 8, when OpenAI announced that an unreleased model had cracked a version of a famous problem about the Navier-Stokes equations, which describe how fluids move (and which engineers use to design aircraft and forecast weather). Picture a whirlpool that spins faster as it narrows, the way a figure skater speeds up by pulling in her arms. The question is whether the equations can ever let that spin run away to infinite speed. In 2000 it became one of seven Millennium Prize Problems, each carrying a $1 million reward, and only one of the seven had been solved before.

According to OpenAI, roughly 10,000 copies of its model, working together as AI agents, took 88 hours to find such a breakdown in a fluid pushed by a smooth outside force. That version is allowed under the prize rules, although some mathematicians argue that it sidesteps the question most of them had in mind, and the institute that awards the prize has not yet ruled. The proof came with a computer-checked version, which works a little like a spell-checker: it confirms that every word is spelled correctly, though not that you wrote the essay you meant to write. In October, researchers found places where the two versions do not fully match, without showing that the result is wrong.

But focusing on one result risks missing the forest for the trees. Whatever becomes of that proof, many mathematicians say their field will never be the same. A few years ago, chatbots tripped over arithmetic that a student would get right. This year a project called First Proof has been testing AI on unpublished research problems, with expert mathematicians grading the answers. The top performers were able to solve a few in February, six or seven out of ten in June, and nine out of ten in October.

That same month OpenAI released 722 papers claiming advances on hundreds of problems. By mid-October it had withdrawn three and revised 14. It will take years to sort through the papers, though the computer scientist Scott Aaronson has already called it one of the biggest days in mathematical history. Solving one of the harder problems on its own could have made a young mathematician's career.

These AI systems learn from what people have written, from textbooks to libraries of proofs that mathematicians assembled over many years. Because a computer can check whether a proof works, they can also improve by trial and error, and teams of agents can split a problem into pieces and keep trying, rather like a research group that never eats, sleeps, or worries about funding or graduation.

The speed has unsettled even mathematicians who use AI themselves. In September, 25 winners of the Fields Medal, mathematics' highest honor, said in a joint statement that the goals of AI companies and of mathematicians are "severely misaligned". One of them, James Maynard of the University of Oxford, told NPR that it had been hard to extract any human understanding from the fluid proof. Terence Tao of the University of California, Los Angeles, has written in Nature of "proof indigestion", with results arriving faster than people can absorb them. Elsewhere he has argued that a correct result is of limited use to other researchers if they cannot work out why it is true. It is like copying the answer from the back of the textbook: you have the number without the method.

The same forces are at work across science, where models comb DNA sequences, molecular structures and telescope images for useful patterns. Mathematics may have felt the change first, but it is far from alone. Astronomy, which is largely data analysis applied to the sky, is also being upended. In a Science feature titled "The last astronomers", Joshua Sokol described how the Harvard astronomer Alyssa Goodman had spent years trying to separate the motion of a galaxy's spiral arms from other effects in her data. ChatGPT resolved the problem within minutes. Her colleague Cecilia Garraffo had a different experience. When she asked AI systems earlier this year to work out a hard problem about a spinning black hole, Claude produced an impressive-looking paper that, on close reading, repackaged existing ideas. Her expertise let her see what was missing.

Someone at my son's stage has not yet had the chance to acquire that expertise. Trainees and even experienced researchers are now handing more and more of their thinking to AI. Minas Karamanis, a cosmologist at the University of California, Berkeley, has argued that a student who lets the machine do the thinking has "forfeited the experience that the time was supposed to give you".

Two days after OpenAI's announcement, the mathematician Ken Ono told about 150 students, junior researchers and faculty at the University of California, Berkeley, that they might graduate into a profession that no longer exists in the form they expected. A talk scheduled for 50 minutes ran past two hours as students pressed him with frustrated and sometimes angry questions.

There are fewer chances to gain experience on the job, too. A Stanford study updated in August found that employment of 22- to 25-year-olds in the United States in the jobs most exposed to AI was 19% below where it would have been had it kept pace with their peers in other work, mostly because fewer were hired. Older workers showed no such gap. The authors caution that the pattern does not prove AI is the cause. Still, a ladder with fewer bottom rungs raises a question about where tomorrow's senior scientists will come from.

Some mathematicians are beginning to rethink training, too. In September, 24 of them, meeting on doctoral education, recommended judging students less by their written work and more by their ability to explain and defend it in person. Their rule of thumb is that AI should "accelerate understanding, not bypass understanding", and they advise even keen users to learn deliberately how to do things by hand.

No one can predict just how deeply AI will change scientific professions and what they will look like in the future. I want my son to keep exploring science while learning enough statistics and computing to question a convincing answer. Broad knowledge can help him spot connections and choose problems worth solving, which still depend largely on people, and flexibility will matter as the tools change. For now, I hope he keeps his curiosity and takes the time to understand the difficult parts, even when a machine offers a shortcut.

And if all else fails, Geoffrey Hinton, the Nobel laureate often called the godfather of AI, has another suggestion. He expects machines to take a long time to master physical work. Asked last year about careers in the midst of superintelligent machines, he offered simple advice: "Train to be a plumber."