REVIEW 3 major objections 4 minor 25 references
The crisis of AI-generated mathematics
T0 review · 3 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash
Pith's one-line read This essay argues that AI-generated proofs, however correct, strip mathematics of its value because the worth of mathematics lies in the human practice of writing proofs, and it calls on the community to organize total opposition.
desk verdict A forceful polemic with one genuinely new idea (co-ownership), undermined by an unresolved tension between its core premise and that proposal. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing distinction is between 'artificial mathematics' and 'natural mathematics.' The mechanism is the research paper understood as a certificate: writing a proof is the deepest form of mathematical practice, and a completed paper certifies that a human achieved complete understanding. When AI generates the proof, the certificate loses its meaning, so the community's ways of evaluating mathematicians, refereeing, and assigning authority all break down. The paper's proposed repair is co-ownership, where showing authoritative understanding of a result, rather than having written it first, becomes the basis for claiming authorship.
What would settle it
Controlled comparisons of two groups of graduate students, one that learns a nontrivial theorem by reconstructing and writing its proof and one that studies a machine-generated proof of the same theorem, could settle the premise: if the reading-only group shows equal understanding, creativity, and retention, the claim that writing is indispensable is falsified. A more direct test would be a mathematical community that relies only on AI-generated proofs yet maintains deep understanding and cohesive practice over several years.
Extended reading notes
Core claim
The paper's central claim is that AI-generated mathematics, even when correct, has no mathematical value for humans because it bypasses the practice of writing and understanding proofs, which is where mathematical value is created. The author argues that this decoupling of practice from measurement will end mathematics as a meaningful human activity: papers no longer certify understanding, the journal and refereeing system drowns in AI output, and human mathematicians become mere messengers between machines. The positive discovery is a proposed alternative, 'natural mathematics,' organized around publicly demonstrated authoritative understanding, including a co-ownership model in which anyone who can explain a paper in full may claim co-ownership, and institutional policies that fund, publish, and reward AI-free work.
Load-bearing premise
The load-bearing premise is that the value of mathematics lives in the human activity of writing and understanding proofs, not in the theorems themselves; if a correct proof produced without human understanding still counts as mathematical value, the argument collapses.
Editorial extensions
If this is right
- If the central claim is right, a correct AI-proved theorem is not a mathematical achievement in the human sense, so prioritizing it misdirects the field.
- The journal system will be flooded faster than humans can referee, forcing journals to either automate verification or shift to certifying human understanding.
- Mathematicians who decline AI will need new signals of value, and departments can provide them through hiring lines, tenure criteria, and support for seminars and conferences.
- A co-ownership model could replace traditional authorship, with multiple humans claiming shared ownership by demonstrating full understanding.
- Institutions that control grants, awards, and public framing can make 'natural mathematics' viable by concentrating resources on AI-free work.
Reading between the lines
- If the author's premise is accepted, an immediate testable prediction follows: mathematicians who learn results only by reading AI-generated proofs will retain less durable, transferable understanding than those who write proofs themselves; this could be measured in a controlled graduate-student study.
- The co-ownership proposal could be piloted by a small journal before being adopted field-wide, much as new authorship norms have historically been tested in specialty communities.
- The argument extends beyond mathematics: any intellectual profession whose output is evaluated by written artifacts, such as theoretical law or philosophy, faces the same collapse of certification if artifacts can be produced without understanding.
- The essay's political call implies that the future of mathematical practice is a matter of collective choice, not technological necessity; this could be examined historically by looking at professions that successfully resisted automation through norms.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This math.HO essay argues that AI-generated mathematics poses an existential threat to mathematics as a human practice. The author claims that the value of mathematics lies not in the theorems produced but in the human acts of discovering, writing, and understanding proofs. On this basis, the essay recommends total opposition to the use of AI in mathematics and proposes a set of actions for individuals, departments, journals, and institutions, culminating in a 'natural mathematics' program whose flagship idea is replacing traditional authorship with co-ownership of works by anyone who demonstrates 'authoritative understanding.' The essay draws on reported recent events involving the AI systems Danus, Fable, and OpenAI models to motivate the crisis, and it explicitly frames the piece as a political and ethical call to action rather than a technical proof.
Significance. If the essay's normative premise is accepted, its call to organize against AI use in mathematics is a consequential intervention in an ongoing public debate, and its concrete proposals for departments and journals are at least actionable discussion points. The essay is clearly written and engages directly with named sources such as Tao, Tsimerman, and the Leiden Declaration. However, the paper's central claim is a value judgment rather than an empirical or mathematical theorem, and the load-bearing constructive proposal in Section 2 is in tension with Section 1's premise. The reported AI events are extraordinary and are cited through non-peer-reviewed or secondary sources, so the evidentiary basis is thinner than the rhetoric suggests. The manuscript therefore needs substantial clarification before its central argument can be considered coherent.
major comments (3)
- [Section 2, co-ownership proposal] The co-ownership proposal directly contradicts the premise established in Section 1. Section 1 states that papers produced without 'the human understanding that comes from writing them ourselves' have little mathematical value, and that authorship cannot be replaced by close reading or digestion. Section 2 then entitles any mathematician who demonstrates 'authoritative understanding of a work – the type you would expect of an author today' to claim co-ownership, even after publication, with 'explaining an entire paper in full detail' given as an example of the required social work. If such understanding can be achieved by reading, teaching, and explaining an AI-generated paper, then human mathematical practice survives without human authorship, and the central claim that AI necessarily ends mathematical practice fails. If such understanding cannot be achieved without independently writing the proof, then the proposal reduces to requiring the very practice the essay says AI eliminates. The manuscript never defines 'authoritative understanding' operationally, so the essay's only concrete institutional alternative cannot rescue the central argument.
- [Section 1, opening narrative and evidence base] The crisis narrative depends on reported events whose status is not disclosed: the Danus proof in Cheng-Liu-Gao [5], the Fable disproof of the Jacobian Conjecture [19], the OpenAI ten problems [16], and the claimed model behaviors [17, 25] are presented as established facts, but the cited sources are preprints, blog posts, press articles, and unreviewed web pages. Because the essay's urgency rests on these events, the author should either explicitly mark them as reported-but-unverified developments or provide a transparent assessment of their reliability. This is not a mathematical correctness issue, but it is a load-bearing evidentiary point for a piece whose conclusions are meant to guide professional policy.
- [Section 1, 'practice' dichotomy] The argument rests on an unexamined dichotomy between writing papers and merely reading them. The essay asserts that the work one has 'written up' is understood best, and uses this to reject Tao's 'digestion' paradigm. Yet Section 2 celebrates 'explaining an entire paper in full detail' as valuable social mathematics. The two views can be reconciled only if explaining a paper to a skeptical audience is not the same as 'digestion,' but the essay does not explain where the line falls. This unresolved distinction matters because it determines whether the 'natural mathematics' program is genuinely a new practice or simply a relabeling of the activity the essay dismisses as messenger work.
minor comments (4)
- [References, [10]] The reference for Hudon and the 'AI psychosis' article appears malformed; the author's name and the article title seem merged, and the bibliographic entry should be corrected.
- [Abstract and Section 1] The terms 'artificial mathematics' and 'natural mathematics' are introduced without definitions; since the entire essay revolves around this opposition, a brief working definition of each would help readers evaluate the proposals.
- [Section 2, individual actions] The essay proposes 'AI vegetarian' and 'AI vegan' identities and says that defining them is 'interesting and important work,' but it does not offer even provisional definitions. A few examples of what each category would and would not permit would make the proposal less abstract.
- [Section 1, four numbered points] The four numbered observations about seminar invitations, graduate student independence, hard problems, and competition are offered as evidence that solving more problems is not the goal of mathematics, but the connection between each observation and the conclusion is left implicit; adding one sentence of explanation for each point would strengthen the argument.
Circularity Check
No circularity: the essay argues from explicit premises and external sources; no prediction reduces to an input.
full rationale
This is an opinion essay, not a derivation. The central claim that AI tools end mathematical practice rests on the stated premise that mathematics' value lies in the human act of writing and understanding proofs ('If papers are produced without the human understanding that comes from writing them ourselves, they have little mathematical value to us'), and the conclusion follows by applying that premise to AI-generated papers. A premise is not a circularly defined output; the essay makes no quantitative predictions, fits no parameters, and invokes no self-citation chain. The cited works by Tao, Tsimerman, Thurston, and the AI companies are external sources, and the paper engages them as interlocutors rather than as authorities that define the conclusion. The paper also includes its own acknowledged limitation: 'Some of these ideas will be good, and others will be bad; they're new, and they need to be tested.' The Section 2 co-ownership proposal may be in tension with the Section 1 claim that reading cannot replace authorship, but that is an internal inconsistency, not circular reasoning: the proposal does not define or derive the central claim from itself. No circular step can be exhibited.
Assumptions & free parameters
assumptions (4)
- domain assumption Mathematical practice (doing math) is an end in itself, not just problem-solving.
- ad hoc to paper Papers certify that the author practiced and understood mathematics; AI-generated papers do not.
- domain assumption Mathematicians can collectively resist AI through departmental and journal policies.
- domain assumption The cited AI events (Danus, Fable, OpenAI) are accurately reported by the referenced articles.
Cite this review
Pith. "Pith review of The crisis of AI-generated mathematics." pith.science (2026). https://pith.science/paper/GCOIXLOF
@misc{pith2026260802859,
author = {Pith},
title = {Pith review of: The crisis of AI-generated mathematics},
year = {2026},
howpublished = {\url{https://pith.science/paper/GCOIXLOF}},
note = {Machine review of arXiv:2608.02859}
}
read the original abstract
In this essay, I present the case for total opposition to the use of artificial intelligence in mathematics. I offer proposals for how individuals, departments, journals, and institutions can act in concert to make sure that mathematics survives the coming crisis.
Reference graph
Works this paper leans on
- [5]
- [19]
-
[16]
OpenAI. Ten advances in mathematics and theoretical computer science.https:// openai.com/index/ten-advances-in-mathematics/
-
[1]
Terence Tao on how AI is changing mathematics.https://x.com/OpenAI/status/ 2060451757818601808. 6
-
[2]
J. Alper. Embracing AI and formalization: experimenting with tomorrow’s mathematical tools.Bull. Amer. Math. Soc. (N.S.), 63(2):177–197, 2026
work page 2026
-
[3]
M. Bolan, J. Breitner, J. Brox, N. Carlini, M. Carneiro, F. van Doorn, M. Dvorak, A. Goens, A. Hill, H. Husum, H. I. Mejia, Z. A. Kocsis, B. L. Floch, A. L. Bar- on, L. Luccioli, D. McNeil, A. Meiburg, P. Monticone, P. P. Nielsen, E. O. Osazuwa, G. Paolini, M. Petracci, B. Reinke, D. Renshaw, M. Rossel, C. Roux, J. Scanvic, S. Srini- vas, A. R. Tadipatri,...
arXiv 2025
-
[4]
R. Cheng. Tangent classes for matroid building sets, 2026.https://arxiv.org/abs/ 2606.22650
arXiv 2026
-
[6]
A Taxonomy of Omnicidal Futures Involving Artificial Intelligence
A. Critch and J. Tsimerman. A taxonomy of omnicidal futures involving artificial intelli- gence.https://arxiv.org/abs/2507.09369, 2025
work page Pith review arXiv 2025
Show all 25 references
-
[7]
First Proof.https://1stproof.org/
-
[8]
Hartnett
K. Hartnett. How Terry Tao Became an Evangelist for AI in Math.https: //www.quantamagazine.org/how-terry-tao-became-an-evangelist-for-ai-in-math- 20260608/
-
[9]
Hartnett
K. Hartnett. Sometimes being first means seeing the end before anyone else. https://www.quantamagazine.org/jacob-tsimerman-wins-2026-fields-medal-for- andre-oort-conjecture-proof-20260723/
2026
-
[10]
Hudon and S
A. Hudon and S. E. Delusional experiences emerging from ai chatbot interactions or ”ai psychosis”. JMIR Ment Health. 2025 Dec 3. doi: 10.2196/85799
2025 doi
-
[11]
Leiden Declaration on Artificial Intelligence and Mathematics.https:// leidendeclaration.ai/
-
[12]
J. Liu, G. Gao, Z. Sun, B. Wu, S. Liu, J. Jiang, H. Ju, L. Chen, R. Cheng, X. Zhang, and B. Dong. Danus: Orchestrating mathematical reasoning agents with fact-graph memory. https://arxiv.org/abs/2607.06447, 2026
2026 arXiv
-
[13]
Luong and V
T. Luong and V. Mirrokni. Accelerating mathematical and scientific discovery with Gem- ini Deep Think.https://deepmind.google/blog/accelerating-mathematical-and- scientific-discovery-with-gemini-deep-think/
-
[14]
Math Inc.https://www.math.inc/vision
-
[15]
An OpenAI model has disproved a central conjecture in discrete geometry.https: //openai.com/index/model-disproves-discrete-geometry-conjecture/
OpenAI. An OpenAI model has disproved a central conjecture in discrete geometry.https: //openai.com/index/model-disproves-discrete-geometry-conjecture/
-
[17]
S. Ornes. The AI was fed sloppy code. it turned into something evil.https: //www.quantamagazine.org/the-ai-was-fed-sloppy-code-it-turned-into- something-evil-20250813/
-
[18]
S. Roberts. As A.I. Makes Strides in Mathematics, Mathematicians Urge Caution.https: //www.nytimes.com/2026/06/02/science/ai-mathematics-leiden-declaration.html
2026
-
[20]
Tao.https://mathstodon.xyz/@tao/116477351524980995
T. Tao.https://mathstodon.xyz/@tao/116477351524980995
-
[21]
T. Tao. A digestion of the Jacobian conjecture counterexample.https: //terrytao.wordpress.com/2026/07/21/a-digestion-of-the-jacobian-conjecture- counterexample/
2026
-
[22]
T. Tao. Mathematics in the Age of AI.https://teorth.github.io/tao-web/slides/age- of-ai-icm-2026.pdf
2026
-
[23]
W. P. Thurston. On proof and progress in mathematics, 1994.https://arxiv.org/abs/ math/9404236
1994 arXiv
-
[24]
Venkatesh
A. Venkatesh. Some thoughts on automation and mathematical research.Bull. Amer. Math. Soc. (N.S.), 61(2):203–210, 2024
2024
-
[25]
S. Witt. Inside OpenAI’s hack of Hugging Face. The New Yorker,https:// www.newyorker.com/news/the-lede/inside-openai-hack-of-hugging-face. Email address:mweinreich@math.harvard.edu 7
Reviewed August 15, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.