An unreleased OpenAI model has produced more than 700 research mathematics manuscripts, and the people who check mathematics for a living are now arguing about how to handle them. Their debate has a line about education, and it arrives as New Zealand schools bed in a new maths curriculum.
OpenAI has published the collection on GitHub. The repository counts 719 manuscripts in 372 families, drawn from about 4,000 problems the model was set. Some tech podcasts have taken to calling it the AI mathpocalypse.
Around 42% of the headline results have been formalised in Lean, software that checks every step of a proof. The repository says not every result has a formal proof yet and some of the unformalised ones may have problems. The day after launch, OpenAI withdrew three manuscripts because a sign error broke the argument in one paper and two that relied on it, and revised 14 others.
Mathematicians set ground rules
The International Mathematical Union has endorsed the Leiden Declaration on Artificial Intelligence and Mathematics, published in June. It says responsibility for the correctness of AI-assisted research stays with the human authors. It also says press releases and blog posts can’t replace peer review.
Terence Tao, a Fields medallist at UCLA, gave a public lecture on AI at the 2026 International Congress of Mathematicians and has since posted it as an essay. Without changes to policy and culture, Tao warns, AI-generated proofs could pile up faster than mathematicians can verify them, and verified proofs faster than anyone can write them up readably.
He points to the First Proof project, an independent test using research problems whose solutions had never been posted online. In its second batch, seven of the 10 problems got at least one passing grade from at least one AI system, at a compute cost of tens to hundreds of dollars per problem.
Tao then turns to teaching. In education and in training young mathematicians, he writes, AI tools should be restricted quite tightly, because “the goal of training a mathematician is not achieved by producing correct homework.”
Where New Zealand schools stand
The Ministry of Education’s generative AI guidance says human teachers must remain central in students’ education and take responsibility for their learning. Schools must have a policy on acceptable use of GenAI to guide NCEA assessment, and GenAI isn’t permitted in NCEA external assessment.
ERO’s report Ready or not, on how schools are responding to AI, found 72% of secondary teachers report negative effects of student AI use on critical thinking. Among primary teachers, 51% said they used AI to some extent in mathematics and statistics, against 73% in English. One primary student told ERO that heavy users don’t learn to do the maths when they can just ask AI.
The achievement picture comes from the Curriculum Insights and Progress Study. In 2025, 24% of Year 8 students were at or above the expected curriculum level in maths, against a 22% baseline in 2023. The Government’s target is 80% by 2030, and the benchmarks are still provisional, set against the 2023 draft curriculum.
Further up the school, NCEA has its numeracy co-requisite. In the first assessment event of 2026, 51.7% of the 61,846 students who sat numeracy achieved it, and around half of all co-requisite participants were in Year 10. Across 2025, 63% of 94,098 students achieved numeracy, counting each student’s last result.
The curriculum timeline
The refreshed Mathematics and Statistics learning area is already required for Years 0 to 10. The Ministry describes it as knowledge-rich, setting out what students should know and be able to do at each year level. The new Year 11 curriculum follows in 2028, Year 12 in 2029 and Year 13 in 2030.
Budget 2026 included a $131 million literacy and maths package under the Foundational Learning Action Plan, with hands-on maths teaching kits among the spending, according to the Government’s target report. By mid-2026, around two-thirds of Year 0 to 10 state and state-integrated schools and teachers had taken part in at least one day of Ministry-funded maths PLD.
On the research side, the First Proof project plans further batches, and OpenAI says it will keep adding Lean formalisations to its collection as it gets them.
