Rhombic tiling, permutation plot, subset gamegraph, permutahedron, hyperplane arrangement, structure diagram, three-dimensional heap, Cayley diagram, braid graph, and L-tromino tiling.

AI and mathematics

A weekly digest of AI-related developments in mathematical research, teaching, and learning.

Weekly digest, 25–31 October 2026

Mathematics advocacy · Institutional position

The mathematical sciences in an AI-enabled world: AcadMathSci shares new preprint

Academy for the Mathematical Sciences · 28 October 2026

In announcing a new preprint, the Academy argues that human expertise remains essential for choosing questions, assessing AI-generated results, developing understanding, and translating discoveries into use. Its announcement identifies education and skills, independent verification, access to AI capabilities, research assessment, and mathematical knowledge infrastructure as community priorities. The preprint is presented as a contribution to discussion and a collective response.

Research and formal verification

Formalizing the enormous theorem

Axiom Math · 26 October 2026

Axiom announces a Lean formalization of the classification of finite simple groups, produced by AI agents coordinated by mathematicians and engineers. Its account emphasizes the work of matching definitions, hypotheses, and dependencies across a large body of existing mathematics, and investigating suspected errors in the literature. The company links the proof and dependency graph for scrutiny. This is a report of Axiom’s announcement, not an independent audit of the formalization or its correspondence with the intended theorem.

Research practice and community governance

Working with academic communities to build scientific understanding

OpenAI · 26 October 2026

OpenAI describes experiments with different ways of releasing AI-generated research and involving academic collaborators. It treats raw AI manuscripts as research data and proposes expert help with checking, interpretation, and exposition. The company states that participation is voluntary and unpaid, collaborators may discuss and publish freely, and human contributions should receive credit. It acknowledges that selecting collaborators creates unequal access to information. These are the company’s stated practices and commitments; the announcement does not establish which release model best serves mathematics.

Research practice and community governance

Disadvantages of designating mathematicians for AI output releases

Advisory Group on Mathematics and Artificial Intelligence, guest post on Terence Tao’s blog · 25 October 2026

AGMAI argues against AI companies selecting mathematicians to lead the interpretation of released results. While acknowledging the value of early expert scrutiny, it identifies risks involving unequal access, community autonomy, selection bias, and borrowed reputation. It recommends transparent releases on independent repositories and support for workshops and working groups allocated through independent nonprofit institutions. Read alongside OpenAI’s announcement, this position clarifies a substantive disagreement about who should organize and govern the work of understanding AI output.

Research access and governance

Mathematicians warn AI firms are keeping vital research in the field secret

James Tapper · The Observer · 31 October 2026

Reporting on a Royal Society meeting, Tapper describes mathematicians’ concerns about private control of advanced AI tools and access to mathematical knowledge. Martin Bridson distinguishes immediate restrictions on research tools from the longer-term risk of knowledge being withheld or placed behind paywalls. Ana Caraiani raises unequal institutional access, while Jon Keating emphasizes that claims must be presented fully enough for the community to check and understand them. The report adds questions about access and accountability to this issue’s discussion of research releases; the concerns it records do not establish that firms are already hoarding fundamental results behind paywalls.

Teaching and assessment

Can artificial intelligence (AI) grade like a teacher? Evaluating skill-based open-ended questions in mathematics

Hasan Fehmi Özdemir · Sage Open · 28 October 2026

This open-access study compares three teachers and four AI applications grading ten open-ended mathematics questions answered by 196 seventh- and eighth-grade students in Türkiye. The tools differed substantially: ChatGPT and Copilot tended to award higher scores, and agreement across individual raters was only moderate. The distinction between ranking students similarly and awarding the same marks is particularly useful for assessment design. Results concern specific consumer tools, one sample, and a single rubric whose scoring discretion may itself contribute to disagreement; they do not establish reliable autonomous grading.

Student formation and teaching

AI, mathematics, education, and a public goods problem

Gireeja Ranade, guest post on Terence Tao’s blog · 29 October 2026

Ranade argues that maintaining human mathematical understanding is a shared responsibility that cannot be reduced to producing answers efficiently. Drawing on conversations with students, she connects motivation and career prospects with the sustained effort needed to develop expertise. She calls for demanding, supported learning, recognition of exposition and understanding, and resources for teachers and individual tutoring. This is an informed argument about educational priorities, not a measured forecast of jobs or evidence that a particular intervention improves learning.

Mathematical culture and student formation

A mathematician reads the AI future

Steven Strogatz · The Atlantic · 25 October 2026 · Gift link

Strogatz explores the difference between obtaining mathematical results and the human experience of discovering and understanding them. Through accounts of colleagues and students at Cornell, he contrasts enthusiasm for AI-assisted research with concern about training opportunities, recognition, and the joy of discovery. The essay presents mathematics as a creative and communal activity while leaving open how that value can coexist with increasingly capable machines. It is a personal reflection on mathematical culture, not a study of learning or career outcomes.

Video · Mathematical understanding

Steven Strogatz on the humanity of math

Grant Sanderson · 3b1b Podcast · 27 October 2026

The publisher describes a conversation with Strogatz about what mathematics means when machines can produce answers. Its chapter outline covers advice for students considering mathematics, the future of the profession, understanding, teaching, and doctoral study, alongside a paper-folding exploration. The discussion offers an optimistic perspective on continuing to learn and do mathematics, complementing the concerns raised in Strogatz’s essay.

Video · Evaluating AI research capabilities

First Proof: A nuanced assessment of AI’s ability to solve math problems

Harvard Mathematics Department · 28 October 2026

Harvard’s description introduces Lauren Williams reflecting on her participation in First Proof, a collaborative experiment designed to assess AI systems’ mathematical capabilities and their relationship to human expertise. The video covers what she has learned from the project and its future direction. It offers a researcher’s perspective on evaluating AI in mathematics, with a link to the project’s latest problem batches.

Editorial note. This issue includes ten items published from 25 through 31 October 2026. Quanta, 3Blue1Brown, and Numberphile were reviewed as standing sources. The video entries summarize publisher descriptions and available chapter information. The Atlantic gift link provided free access during review; availability may change. The content is generated by ChatGPT from linked sources and may contain errors or omissions; it is a starting point for review and does not independently verify reported mathematical claims.
Coverage and access. Only freely accessible material is summarized here. Some entries cover a freely available abstract rather than a full article; these are explicitly labeled, and the full article may require a subscription. This collection is selective, not exhaustive. Paywalls and other access restrictions can limit coverage, and sources that cannot be adequately verified are omitted pending review.