In early October OpenAI published more than 700 AI-generated maths preprints on GitHub, covering combinatorics, geometry, number theory, topology and more. It claimed progress on 372 problem families out of roughly 4,000 attempted, including partial advances tied to three of the unsolved Millennium Prize Problems. Mathematicians are, in the words of one headline, in mourning.
Business owners should care less about the mourning and more about one number. At the time of reporting, only 300 of the 719 manuscripts, or 42%, had been formally verified. The machine producing expertise has outrun the system built to check it. That is the real story, and it is coming to law firms, actuaries, engineers and banks.
Scarce expertise just became scalable
Open problems in mathematics are the scarcest intellectual commodity there is. Some consumed careers. OpenAI chewed through hundreds of them using a fraction of the compute behind its earlier headline breakthrough.
Paul Baier, writing in Forbes, puts the commercial lesson bluntly: the system worked “in weeks, creating new knowledge at an unmatched pace,” and every company depends on some form of cognitive advantage. If the hardest expertise on earth can be mass-produced, the premium on your firm’s specialists is not safe by default.
The human cost is already visible. NYU mathematician Tristan Buckmaster said he knew of three people whose entire research programmes had been obliterated by the release. Swap “research programme” for “billable niche” and the parallel writes itself.
The bottleneck moved to checking
The sharper insight comes from the critics. Edward Witten of the Institute for Advanced Study called the work “often impressive… but the presentation a mess,” while UC Berkeley’s Olga Holtz drew a line between a claimed result, a verified one and one properly absorbed into the field: “those are not the same.”
Fields medallist Terence Tao argues that “proof scarcity” is no longer the binding constraint, because AI now generates results faster than mathematicians can digest them. Philosophers Silvia De Toffoli and Eamon Duede put it more pointedly about an earlier OpenAI result: “What OpenAI has given us is an answer. But it is not clear that they have delivered a fruitful solution.”
An answer you cannot understand is an answer you cannot build on, defend in court or explain to a regulator. Business already pays for this. Developers using AI coding tools spend an average 11.4 hours a week reviewing AI-generated work, according to a survey of 2,847 users. The productivity dividend is real, but a big slice of it is spent on verification.
Your bank’s risk model has the same problem
Tao’s most commercially important warning is easy to miss. AI-generated mathematical models are already used, to some extent, to determine bank loan eligibility and measure market risk, without matching human understanding of how they work.
That is a governance blind spot, not an academic curiosity. NZ banks, insurers and the Reserve Bank and FMA should be asking now who in the building can actually explain the model, rather than finding out after one fails quietly. Boards signing off on AI-assisted credit, pricing or valuation work should demand the same answer.
New Zealand is more exposed than it looks
This is not someone else’s problem. Back in 2024, Treasury warned that AI’s outsized impact on higher-skilled tasks meant higher-skilled economies like New Zealand faced greater employment exposure, not less, and that slow technology diffusion could blunt the upside.
The upside is large. MBIE’s 2025 AI strategy cited modelling that generative AI could add up to $75 billion to the economy by 2038. Yet the same document flagged that 68% of SMEs had no plans to even evaluate AI. A high-skill service economy that refuses to look at the technology aimed squarely at high-skill services is not playing it safe. It is waiting to be priced out.
Redeploy the experts, don’t fire them
The lazy conclusion is that expertise is finished. The evidence says otherwise. A 2025 Treasury briefing noted high-skilled workers face substitution risk but are more likely to be complemented by AI than simply replaced, with little consensus on net job effects.
The maths episode shows where the value moves. When answers become cheap, judgement becomes expensive. Firms that shift their senior lawyers, actuaries and engineers from producing first drafts to checking, stress-testing and applying AI output will capture the gains. Firms that let output pile up faster than anyone can verify it will end up like OpenAI’s GitHub repository, with impressive volume and 58% of it unchecked.
Mathematicians will need years to sort through October’s drop. Most New Zealand businesses will not get years. The next release will land in their own field, and the question will be whether anyone on staff can tell which answers are right.
Sources
- Nature: OpenAI posts 700 maths preprints online, mathematicians are up in arms (2026-10-09)
- BusinessDesk: AI solved one maths problem and everyone freaked out. It just cracked hundreds more (2026-10-07)
- ODT: Mathematicians in mourning as AI cracks their lives’ work (2026-10-23)
- The Verge: ‘Pure insanity’, mathematicians will need years to make sense of OpenAI’s latest drop (2026-10-09)
- Nature: Can maths survive AI? Fields medallist Terence Tao calls for a ‘Maths 2.0’ world (2026-10-13)
- Treasury: Official Information Act response 20250500, future of AI (2025-08-13)
Join the discussion
Add useful context, ask a good question, or challenge an idea — keep it specific and respectful.