News

What AI Math Breakthroughs Mean for Expert Work

AI systems are now producing what researchers describe as publishable, PhD-level mathematics — and a recent IEEE Spectrum investigation into artificial intelligence in mathematics is drawing a sharper line than most coverage has: this is not a future risk, it is a present structural shift in how expert knowledge work gets done.

Google DeepMind's Aletheia system generated a research paper on eigenweights without human intervention, and separately resolved open problems from the Erdős Conjectures database while operating with minimal human guidance. According to MarkTechPost's coverage of DeepMind's published findings, Aletheia also scored 95.1% on IMO-Proof Bench Advanced, a benchmark of Olympiad-level proofs, up from a previous record of 65.7%. Meanwhile, models from multiple labs have reached gold-medal standards at the International Mathematical Olympiad.

Mathematicians are now openly split on what this means for the profession. Some warn of a future where human experts become, as one researcher described to IEEE Spectrum, "priests to oracles" — intermediaries who relay machine output rather than generate ideas. Terence Tao, the UCLA mathematician and Fields Medal winner, offers a more structured alternative he calls "big mathematics": large-scale, decentralized collaborations between humans and machines, where people claim the creative and judgment-intensive work while AI handles technical volume. In a March 2026 post reported by OpenAI Academy, Tao stated that AI has become "ready for primetime" in math and theoretical physics because, in his framing, it saves more time than it wastes — and he now uses it to compress literature searches from weeks to minutes and to test whether approaches are worth pursuing before committing full effort.

The skill Tao specifically called out as growing in importance is formal verification — the ability to translate a fuzzy mathematical argument into a machine-checkable form using proof assistants such as Lean. He noted that AI can generate polished-looking arguments that hide flawed reasoning, and that verification tools keep AI "honest." The same IEEE Spectrum piece reports that Math Inc.'s Gauss system formalized Maryna Viazovska's Fields Medal-winning 8-dimensional sphere-packing proof — a task that would typically take expert human mathematicians years — in a matter of days, with the 24-dimensional case completed autonomously in two weeks.

What this means for job seekers

Mathematics is an extreme case, but it is not a special case. The pattern the math field is revealing applies across any knowledge-intensive role: AI absorbs the high-volume, rules-bound output layer fastest. What remains — and what commands a premium — is a cluster of three capacities that Tao's own workflow illustrates.

The first is problem selection: choosing which questions are worth pursuing in the first place. The second is formalization and verification: turning a loosely defined business or research problem into a structure precise enough for a machine to act on, then checking whether the output is actually correct. This is not a passive skill; it requires deep domain knowledge. The third is communication and judgment: defending conclusions, weighing tradeoffs, and translating results for non-expert stakeholders — the human side of the loop that machines cannot close.

For job seekers repositioning in high-skill fields, this is a cleaner frame than the usual "learn to use AI tools" advice. The work shifting to automation is raw output — the first draft, the initial calculation, the search through known methods. The work appreciating in value is the layer above it: defining the problem, verifying the answer, and deciding what to do with it. Reviewing how AI is already reshaping technical interview preparation and the data analyst career path offers concrete examples of how this division is already appearing in hiring criteria across adjacent fields.