You chose maths because it felt safe. Now AI solves the problems that proved you were good. What is really happening to the work and to your career.
This in-depth guide covers everything you need to know about will ai replace mathematicians? what terence tao says. Based on verified income data and real-world case studies from our database of 138 side hustle tactics.
AI already does a large share of the work that used to prove you were good at maths: solving hard problems, checking proofs and grinding through calculations. Research mathematicians are still needed, in fewer roles and with a different job. In 2026 AI systems scored full marks at the International Mathematical Olympiad and settled open Erdős problems, and Terence Tao now warns that AI is using up the supply of good research problems. We found no record of him saying mathematicians are irrelevant. What he says is that the job, the career ladder and the way the field rewards people all have to change, and that mathematicians need to decide how before technology companies decide for them.
The degree you paid for
You were the one who liked the hard question. At school you finished the problem set first, and someone told you maths was the safe choice: every industry needs people who can think in proofs and numbers. So you did the degree. Maybe the loan is still there, a monthly line on your statement next to rent. Maybe you went on to a master's because the good jobs asked for one, or you sat actuarial exams on weekends while friends were out.
Then this year a chatbot your little cousin uses for homework scored a perfect paper at the International Mathematical Olympiad, the contest you once dreamed of qualifying for. You read that a Fields Medal winner left his university for an AI lab. You wonder, quietly, whether the thing you were best at is the thing that is now cheapest.
What AI has done in mathematics so far
The progress has been fast, and it is easy to lose track of the dates. Here is the record, with sources.
| Date | What happened | Source |
|---|
| July 2024 | Google DeepMind's AlphaProof and AlphaGeometry reach silver-medal standard at the IMO: 4 of 6 problems, 28 of 42 points, after experts translated the problems into the Lean proof language and the systems ran for two to three days | Google DeepMind |
| July 2025 | Gemini Deep Think scores 35 of 42 (gold standard), solving 5 of 6 problems in plain English inside the human 4.5-hour limit, graded by the IMO | Google DeepMind |
| July 2025 | OpenAI says an experimental model also scored 35 of 42, graded by former medallists, without official IMO certification | The Decoder |
| January 2026 | Erdős problem #728 is solved with GPT-5.2 Pro and verified in Lean by Harmonic's Aristotle, the first solution with no prior human work | Quanta Magazine |
| May 2026 | A DeepMind team reports its best agent "autonomously resolved 9 of 353 open Erdős problems" for a few hundred dollars per problem | Quanta Magazine |
| May 20, 2026 | An internal OpenAI model finds a counterexample to Erdős's 1946 unit distance conjecture | Quanta Magazine |
| July 2026 | Huawei's and Xiaohongshu's models report full marks at IMO 2026 under official judging; 7 of 666 human contestants got full marks | AFP via The Standard |
| September 2026 | GPT-6.1 Sol scores 100% on FrontierMath Tier 4 (v2), Epoch AI's research-level problem set | Epoch AI benchmark data |
Two things stand out. The first is the speed. In two years AI went from silver with human translators and days of compute, to gold in plain language under exam rules, to a perfect paper. The second is the move from contests into research. The Erdős problems are a list of about 1,200 questions left by Paul Erdős, many open for decades, and the Quanta report from August 2026 counts 565 solved and 652 open, with AI behind a growing number of recent solutions.
FrontierMath tells the same story. Epoch AI built it from several hundred unpublished problems written by mathematicians, from undergraduate level up to research level. In Epoch's own data, the best score on the original Tier 4 research set was 47.9% by Google DeepMind's AI co-mathematician in May 2026. On the newer v2 sets, the best model now scores 93.7% on Tiers 1 to 3 and 100% on Tier 4.
Formal proof is the quiet part of this. Lean is a language in which a computer checks every step of a proof. When GPT-5.2 Pro's proof of #728 contained small errors, Tao wrote on Mathstodon on January 7, 2026 that "the AI tool Aristotle was able to automatically repair these gaps and produce a Lean-verified proof." Checking proofs was one of the jobs that kept postdocs and referees busy. A machine now does part of it.
There is honest counter-evidence. Epoch also runs a set of Erdős problems formalised in Lean, and on that the best models score about 2.9%. Many hard problems are still out of reach.
What Terence Tao actually said
Tao, a Fields Medal winner at UCLA, is the mathematician most people quote on this, and he is often misquoted. A claim circulates that he said AI makes mathematicians irrelevant. We could find no post, talk or interview where he says that. Here is what he has said, in his words and with dates.
On how far AI has come for research, at an IPAM conference in March 2026, OpenAI's write-up quotes him saying AI is "ready for primetime" in maths and theoretical physics because it now "saves more time than it wastes." In September 2024 the same write-up records him calling an earlier model "a mediocre, but not completely incompetent, graduate student."
On how much of the open-problem list AI can take, he wrote on January 14, 2026: "My guess is that this proportion is on the order of 1-2%." He added that "human expert attention will remain a significant bottleneck."
On what this does to the field, his September 8, 2026 thread is the clearest statement of the problem. He wrote that "the collection of good, fruitful open problems is now being mined in a non-renewable fashion, leading to the potential scenario of these problems becoming scarce." In the third post of that thread: "We have now seen that even the rumor of someone working on a problem can trigger a massive amount of AI-powered effort to flatten it before the original research project has time to reach its full potential." And: "The incentives may now be pointing in the direction of no longer sharing any promising research directions with the broader community, which would reverse centuries of traditions of open science and do serious long-term damage to the future of the field."
On the people, his October 6, 2026 thread describes problems "being solved autonomously by AI prompters who have no interest in the broader field itself" and says that "few people are joining the community around the field as a consequence". He ends it with this line about careers: "our community will need to explicitly re-evaluate its criteria for education, publication, and career advancement".
At the International Congress of Mathematicians in Philadelphia in July 2026, the Simons Foundation reports he told the audience: "Let's assume that soon, AI will be able to perform a reasonable fraction of mathematical tasks successfully. If you condition for that, it becomes clear that we now have to think about our goals and our values." He also said: "We set the rules on what's acceptable or not, and we should not let external actors define those for us."
On September 11, 2026 Tao was one of 25 Fields Medallists who signed a declaration he posted on his blog. Its central sentence: "The goals of the AI companies and the goals of the mathematical community are severely misaligned."
One more source gets confused with Tao. The post titled "We're gonna need a lot more mathematicians" on his blog, dated September 24, 2026, is a guest post by the cryptographer Amit Sahai. Sahai argues society will need more mathematically trained people to check what AI produces. That is Sahai's argument, hosted by Tao.
So the accurate summary is this. Tao says AI is now useful, that it is taking the easier open problems at scale, that this is draining the field of good problems and new people, and that careers will be judged differently. He is worried. Nowhere in these posts and talks does he call mathematicians irrelevant, and he keeps telling mathematicians to set the terms themselves.
Question to sit with: if the hardest part of your future job is choosing which problem is worth solving, how much of your training so far has been about choosing, and how much about solving?
Who is already losing ground
The research world gives the earliest signals, and they are not comfortable. Noga Alon, one of the best-known combinatorialists alive, told Quanta he had stopped working on Erdős problems: "Once AI started to solve them, there is no point anymore." The same report says Jacob Tsimerman, on the day he received the 2026 Fields Medal, "announced that he was leaving academia for a job at OpenAI," and that Tao himself "has stepped away from the Erdős problem community to focus on getting work done." Thomas Bloom, who runs the Erdős problems site, warned about non-mathematicians posting "100- to 200-page papers" that no human has read.
Think about what that means if you are a PhD student. Many theses start with an open question your adviser thinks is tractable. If a well-funded team can point an agent at the whole list for a few hundred dollars a problem, the easy and medium questions that used to launch careers are the first to go. Tao's description of "fewer seminars, workshops, collaborations" around AI-solved results is a description of fewer entry points.
Outside academia the numbers are smaller but point the same way. Microsoft Research's 2025 study of 200,000 Copilot conversations scored every US occupation for how much its work overlaps with what AI already does. Mathematicians ranked 10th of 785, with a score of 0.39, and the paper's tables show AI activity touching about 90% of their work activities. Only interpreters, historians, writers and a few sales and service roles scored higher.
Anthropic's June 2026 survey of about 9,700 Claude users found that computer and mathematical occupations make up "roughly 30% of survey respondents" against about 4% of US employment, and that more than a third of respondents think a junior colleague has a better-than-60% chance of losing their job in the next year. The people most fluent in maths are also the heaviest users of the tools, and they are the ones who think the juniors are in trouble.
For recent graduates, the New York Fed's table of outcomes by major, updated in August 2026, shows mathematics majors with 5.8% unemployment and 26.2% underemployment, meaning a quarter work in jobs that do not need a degree. Recent graduates overall sit at about 5.6%. In the same table, accounting (2.6%) and finance (2.8%) majors have lower unemployment than maths majors. Computer science (7.0%) and computer engineering (7.8%) are doing worse.
Actuaries, quants, statisticians and maths teachers
Most people with a maths degree never become research mathematicians. They price risk, model markets, analyse data or teach. Here is where each group stands.
| Job | AI applicability score (rank of 785) | US jobs 2025 | Projected change 2025 to 2035 | Median pay 2025 |
|---|
| Mathematicians | 0.39 (#10) | 2,200 | +1% | $126,710 |
| Data scientists | 0.36 (#19) | 275,600 | +35% | $120,230 |
| Statisticians | 0.32 (#71) | 31,300 | +11% | $105,650 |
| Operations research analysts | 0.31 (#85) | 113,100 | +12% | $88,940 |
| High school teachers (all subjects) | 0.18 (#299) | 1,087,500 | 0% | $72,040 |
| Actuaries | 0.16 (#353) | 31,200 | +9% | $130,000 |
AI scores and ranks are from Microsoft Research's occupation data. Jobs, growth and pay are from the US Bureau of Labor Statistics. All US occupations together are projected to grow about 3% over the same decade, according to the BLS projections released August 27, 2026.
Actuaries. The headline numbers look safe: 9% projected growth and $130,000 median pay. Look closer at the Microsoft data, though. The study separates what people ask AI to help with from what AI actually performs, and actuaries were one of the occupations with the biggest gap toward AI performing the work, at the 76th percentile for AI action against the 42nd for user goals. The exams still gate the profession. In other words, the study saw AI doing actuarial work more than helping actuaries do it, and the routine modelling and report drafting juniors learn on is that kind of work.
Quants. We could not find solid public data on whether hedge funds and trading firms are hiring fewer junior quants because of AI, so we will not invent a number. What is documented: operations research analysts, the closest BLS category, are projected to grow 12%. Goldman Sachs Research reported in August 2025 that unemployment among 20- to 30-year-olds in tech-exposed occupations had risen by almost 3 percentage points since the start of 2025. Quant research is coding, statistics and modelling, the kind of information work the Microsoft study found AI does most often. Financial and investment analysts score 0.28 in that study, rank 118 of 785.
Statisticians and data scientists. These are where official forecasts disagree most with the fear. BLS still projects data scientists to grow 35% over the decade, one of the fastest of any job, and ties computer and mathematical growth to "the continued proliferation of digital tools and AI solutions". The Stanford Digital Economy Lab sees a different picture for the young. Its August 2026 paper, using ADP payroll data on millions of workers, finds that employment of workers aged 22 to 25 in AI-exposed occupations "now stands 19% below" where it would be had it kept pace with less-exposed peers, mostly through reduced hiring. Older workers show no comparable gap. The BLS forecast counts total jobs over ten years. Stanford tracks who is getting hired now. Both can be true, and the second is what decides whether you get the first job.
Maths teachers. High school teachers score low on AI applicability and BLS projects no change in their numbers. The pressure on teachers comes from what students can now do without them. When AI models score full marks at the IMO, every homework set can be answered, and a maths tutor's hourly rate competes with a free app. Postsecondary maths teachers score higher, 0.30 or rank 103 of 785. In June 2026 Tao gave a European Mathematical Society lecture titled "How should students control their AI diet?", which tells you where the worry sits. There is more on this in the teachers post in this series.
Question to sit with: if your employer could get the first draft of your model, your reserve report or your lesson plan from a tool, what is the part of your week you would still be paid for?
Where the forecasts disagree
The official forecasts are calmer than the research news. BLS projects every maths occupation above to hold or grow. The World Economic Forum's Future of Jobs Report 2025 expects 170 million jobs created and 92 million displaced worldwide by 2030, with 41% of employers planning to cut staff where AI can automate tasks. Goldman Sachs' baseline is that AI could displace 6% to 7% of the US workforce if widely adopted, with a range of 3% to 14%.
Three things explain the gap. Projections are built on past trends, and AI capability has moved faster than any trend line this year. They count jobs in total, and a job can survive while the number of people hired into it each year falls. And they treat "mathematician" as a stable title, while Tao is telling you the content of that title is being rewritten.
Tao also warns against anyone who sounds too sure. In a July 2026 interview posted on his site he said: "I'm not sure anyone is capable of any reliable forecasting beyond a year at best, currently."
Question to sit with: are you planning your next five years on a forecast that its own experts say cannot see past one?
What this means for you this month
Keep the maths. Move your weight off the part a machine now does for free, toward the part people still pay a person for: judgement, explanation and trust.
- Use the tools on your own work this week. Give your current model, proof or lesson to a frontier model and see what it gets right. You will learn faster where you still add value than any forecast can tell you.
- Sell explanation. Tao's own fix for the field is more exposition, more talks, more people who can say why a result matters. That skill sells outside academia too. Students and parents still pay for a human who can explain, and online tutoring or a small course turns that into income you control.
- Get paid to grade the machines. AI labs hire people with maths degrees to write and check hard problems. It is uneven work, but it pays, and this guide to AI training jobs covers the rates and the catches.
- Know your number. If your role is the kind the hub post on jobs AI will replace flags, work out how many months of costs you hold in savings before you need to decide anything.
If you work near finance, the financial advisors post and the software engineers post cover the jobs many maths graduates move into.