The Technological Turn in Mathematics

Fenner Tanswell; Silvia De Toffoli

REVIEW 2 major objections 6 minor 3 references

The Technological Turn in Mathematics

T0 review · 2 major / 6 minor · reviewed 2026-07-09 · glm-5.2

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read AI tools reshape what counts as proof and who counts as a mathematician

desk verdict Well-executed survey chapter whose central 'turn vs. acceleration' thesis is undercut by its own analysis showing existing frameworks absorb the new technologies cleanly. read the letter →

arxiv 2607.07162 v1 pith:QCDQNEZP submitted 2026-07-08 math.HO

Silvia De Toffoli , Fenner Tanswell This is my paper · ORCID

classification math.HO MSC 00A3003A0503B35

keywords interactivetheoremproverslargelanguagemodelsmathematicalproofformalverificationsocialepistemologytrustfallibilismcrowdsourcing

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that interactive theorem provers, automated theorem provers, and large language models constitute a genuine technological turn in mathematics, not merely an acceleration of existing practice. The central claim is that these tools reshape three things at once: how mathematical justification works (formal verification supplements but does not replace human simil-proofs, introducing new technological fallibilities), how mathematicians collaborate (large-scale crowdsourced projects become possible where interpersonal trust is no longer required because the machine checks every step), and how epistemic labor is divided between humans and machines (neural AI can generate conjectures and proofs, but introduces alignment problems and hallucination risks that human checking cannot reliably catch). The paper's load-bearing mechanism is the concept of a simil-proof—an argument accepted by a mathematical community as meeting professional standards, which may or may not turn out to be a genuine proof. ITPs and LLMs change the reliability of simil-proofs and the social structures that produce them, but they also introduce new failure modes: software bugs, translation gaps between informal and formal mathematics, and the seductive authority of LLM-generated text that mimics correctness without guaranteeing it.

What carries the argument

The simil-proof concept (an argument conforming to community inferential standards, which is usually but not always a genuine proof), the commitment account of trust (trust as reliance on others to fulfill professional commitments), and the hybrid epistemic network (a collective of human mathematicians, automated tools, and formal libraries whose interaction produces mathematical knowledge that no single node possesses).

What would settle it

If ITPs and LLMs ultimately function only as faster tools within existing mathematical practice—accelerating proof-checking and collaboration without altering the epistemic standards, trust structures, or division of labor that define mathematical knowledge production—then the claim of a genuine 'turn' collapses into an incremental technological adoption story.

Watch

Extended reading notes

Core claim

The paper identifies a dual transformation: ITPs enable trust-free, large-scale mathematical collaboration by mechanically verifying every contribution, while LLMs redistribute epistemic labor between humans and machines but introduce systematic errors that human mathematical training is not equipped to detect. The combination of these technologies creates hybrid epistemic networks—collective human-machine systems where mathematical knowledge may reside in the network as a whole rather than in any individual participant, human or machine.

Load-bearing premise

The paper assumes that existing philosophical frameworks for human mathematics—particularly the simil-proof concept and the interpersonal commitment account of trust—transfer to hybrid human-machine systems without needing fundamentally new epistemic categories. If formal verification changes the nature of mathematical justification itself rather than just its reliability, the distinction between a 'turn' and an 'acceleration' becomes harder to sustain.

Editorial extensions

If this is right

If ITP-mediated collaboration eliminates the need for interpersonal trust, mathematical research could open to contributors who lack institutional credentials or personal reputations, democratizing participation in ways previously impossible.
If LLM-produced proofs contain systematically different errors than human-produced proofs, the mathematical community will need new training methodologies for proof-checking, since existing techniques evolved to catch human-style mistakes.
The hybrid epistemic network model suggests that mathematical knowledge attribution may need to expand beyond individual authors to include software systems, formal libraries, and the collective infrastructure of verification tools.
Autoformalisation—automatically converting informal proofs into formal verified ones—could lower the barrier to formalization enough that it becomes standard practice, but the translation gap between informal intent and formal statement remains an irreducible source of potential error.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If formal verification becomes standard for published proofs, the mathematical literature may bifurcate into verified and unverified tiers, creating a new hierarchy of epistemic status that could disadvantage areas where formalization is technically difficult (e.g., differential geometry, as the paper notes).
The cybersecurity vulnerability of formal mathematics—where Lean's Mathlib could be hacked or corrupted—suggests that mathematical infrastructure may eventually require the same security auditing as financial or medical systems, a cost the mathematical community has never had to bear.
The claim that trust shifts but does not disappear in ITP-mediated collaboration may understate the transformation: if the objects of trust become software maintainers and library moderators rather than mathematical peers, the social fabric of mathematics changes character, not just structure.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

This is a well-written survey chapter for the Blackwell Companion. The main thing to know: it synthesizes the authors' prior work on fallibilism and trust with recent developments in ITPs and LLMs, and the synthesis is genuinely useful as a map of the terrain. The Equational Theories Project case study (Section 4) is the freshest material — it concretely illustrates what large-scale ITP-mediated collaboration looks like and why it differs from the Polymath Projects. The four-benefit analysis of ITPs for collaboration (trust-free verification, searchable libraries, asynchronous access, modularization) is clear and well-organized. The discussion of translation gaps between informal and formal mathematics (Section 3) is also solid, drawing on the right sources. The LLM section is honest about both capabilities and concerns, including environmental costs, deskilling, and cybersecurity — points often missing from such discussions. The soft spot is the central thesis. The paper claims these technologies constitute a 'turn' that 'reshapes' mathematical practice rather than merely accelerating it. But the paper's own arguments work against this. Section 3 shows ITPs are another layer of fallibility mitigation with their own bug profiles — the simil-proof concept extends without modification. Section 5 shows trust doesn't disappear but shifts to moderators and library maintainers — same commitment account, just redirected. The one place where genuinely novel epistemology might be needed — the hybrid epistemic network where no single agent holds the full proof — is explicitly left as an open question. So the completed analyses demonstrate continuity, and the one candidate for rupture is unresolved. The stress-test note lands here: the paper's own arguments are evidence against the 'turn' framing rather than for it. This is a real tension, not a minor one. The reliance on corporate press releases for LLM capability claims (footnote 23) is acknowledged but not fully addressed — Buzzard's own blog post, which the authors cite, is more skeptical than the paper's framing suggests. For a Companion chapter, this does its job. It surveys the landscape, organizes the philosophical questions, and gives readers entry points. The 'hybrid epistemic network' framing, if developed further, could be the genuinely novel contribution, but as presented it's a sketch. Recommend for peer review. The 'turn vs. acceleration' tension should be pressed on, but the chapter is a competent and useful piece of work.

Referee Report

2 major / 6 minor

Summary. This manuscript surveys the philosophical implications of emerging computational technologies—Interactive Theorem Provers (ITPs), Automated Theorem Provers (ATPs), and Large Language Models (LLMs)—for mathematical practice. It covers four main themes: (1) how ITPs affect proof, rigor, and fallibility (Section 3); (2) how ITPs enable new forms of large-scale collaboration and crowdsourcing (Section 4); (3) how trust relationships are reconfigured in technology-assisted mathematics (Section 5); and (4) the potentials and concerns raised by neural AI systems (Section 6). The paper introduces or deploys several conceptual frameworks, including the 'simil-proof' concept (from De Toffoli's prior work), a commitment-based account of trust (from De Toffoli and Tanswell's prior work), and the notion of a 'hybrid epistemic network.' The manuscript is written for a forthcoming Blackwell Companion volume.

Significance. The paper addresses a timely and important topic. The philosophical literature has not yet caught up with the rapid deployment of ITPs and LLMs in mathematical practice, and a synthetic overview is valuable. The concept of a 'hybrid epistemic network' (Section 4) and the analysis of the Equational Theories Project as a case study are genuine contributions, providing a concrete philosophical framework for understanding distributed human-machine mathematical knowledge production. The discussion of the alignment problem for LLMs in mathematics (Section 6, drawing on Tanswell and Berg) is well-motivated. The paper does not ship machine-checked proofs or reproducible code, but it does engage carefully with specific formalization projects (Gonthier's Four Color Theorem verification, Hales's Flyspeck project, the Liquid Tensor Experiment) and concrete crowdsourcing cases (Polymath, ETP), which grounds the philosophical analysis.

major comments (2)

[Section 7 (Conclusion) and Section 5] The central thesis — that these technologies constitute a genuine 'technological turn' rather than mere acceleration of existing practices — is not fully defended against the paper's own analysis. The skeptic's concern is substantive: in Section 3, ITPs are shown to be another layer of fallibility mitigation (with bugs, translation gaps, and conceptual misalignment risks), and the simil-proof concept extends to formalized proofs without structural modification. In Section 5, the authors argue trust 'does not disappear' but shifts to moderators and library maintainers — but if the trust structure is isomorphic (same commitment account, same reliance on community engagement, redirected to different agents), this is continuity, not transformation. The one place where genuinely novel epistemology might be needed — the hybrid epistemic network of Section 4 — is explicitly left as an open ('up
[Section 6] Several claims about LLM capabilities and IMO performance are drawn from non-peer-reviewed corporate sources. The 2025 IMO results (Section 6, p. 15-16) are attributed to company claims without independent verification, and the authors themselves note that Buzzard has raised concerns about testing methodology (footnote 23). The AlphaProof results (Hubert et al. 2025, published in Nature) are more appropriately cited. The paper should more clearly distinguish between peer-reviewed results, preprints, and corporate press releases, and should flag the epistemic status of each claim. This is particularly important for a philosophy of mathematics paper, where the reliability of evidence is itself a central theme.

minor comments (6)

[Section 2] The taxonomy (symbolic AI vs. neural AI) is introduced but the third category from Avigad's taxonomy (symbolic AI for automating mathematical reasoning, item b) is mentioned in Section 1 but not discussed in Section 2 or elsewhere. A brief note on why ATPs receive less attention would help.
[Section 3] The term 'simil-proof' is introduced without sufficient standalone explanation for a companion volume reader who may not be familiar with De Toffoli's prior work. A brief parenthetical definition beyond 'an argument that conforms to the inferential standards of a legitimate mathematical community' would help.
[Section 4, footnote 18] The footnote on the recipe model of proofs and propositional vs. practical knowledge is substantive and arguably belongs in the main text, as it bears directly on the epistemological question of what kind of knowledge the hybrid network possesses.
[Section 5] The distinction between 'mere reliance' on computers and interpersonal 'trust' is central to the argument but is asserted rather than fully defended. The authors note that other accounts of trust (Coliva, Nguyen) would classify computer reliance as trust, but do not explain why their preferred account should be privileged here. More defense is needed.
[Section 6] The 'vibe formalisation' example (Alexeev and Mixon 2025) is interesting but the epistemological implications are underexplored. If the human authors cannot read the Lean code, what is the epistemic status of the verification? This connects to the translation problem of Section 3 but is not linked there.
[References] Several entries have forthcoming or 2026 dates that may need updating before publication (e.g., De Toffoli 2026, Tacca 2026, Tanswell and Berg 2026). The Bolan et al. (draft) reference for the ETP should be updated if a version becomes available.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: philosophical survey paper with self-contained argumentation

full rationale

This is a philosophy of mathematics survey chapter, not a derivation-chain paper. There are no fitted parameters, no predictions from first principles, and no equations that could reduce to inputs by construction. The authors do cite their own prior work extensively (De Toffoli 2021a on simil-proofs, De Toffoli and Tanswell 2025 on trust), but these citations provide conceptual frameworks that are applied to new case studies (the Equational Theories Project, AlphaProof, FunSearch, etc.) rather than serving as load-bearing premises that make the central thesis true by definition. The central claim—that ITPs, ATPs, and LLMs constitute a 'technological turn'—is supported by independent case studies and examples drawn from external sources (Gonthier 2008, Hales et al. 2017, Scholze's Liquid Tensor Experiment, the Polymath Projects, DeepMind's AlphaProof). The self-citations function as philosophical scaffolding whose applicability is argued for, not assumed. The one place where a genuinely novel epistemological category is needed (the hybrid epistemic network of Section 4) is explicitly left as an open question rather than being resolved by fiat or self-citation. No step in the paper's argument reduces to its inputs by construction, and the case studies are externally verifiable. This is a normal, honest philosophical analysis with no circularity.

Assumptions & free parameters 0 free parameters · 5 assumptions · 2 invented entities

The paper is a philosophy chapter, so its 'axioms' are philosophical frameworks and domain assumptions rather than mathematical axioms. The most load-bearing is the ad-hoc framing claim that current technological changes constitute a 'turn'—this is an interpretive judgment that structures the entire chapter but is not independently derived. The simil-proof concept has the strongest independent grounding, supported by historical cases of erroneous proofs. The hybrid epistemic network is the most speculative entity, extending an existing framework (social machine) without yet demonstrating clear predictive advantages.

assumptions (5)

domain assumption The Standard View: genuine proofs are arguments that can be formalized in an appropriate formal system.
Invoked in Section 3 as the baseline against which ITP-based verification is discussed. The paper does not defend this view but builds on it.
domain assumption Mathematicians are fallible: they sometimes announce proofs they do not have, and sometimes erroneous arguments are mistaken for proofs.
Stated in Section 3 as motivation for the simil-proof concept and for the need for ITPs. Empirically well-supported but treated as an axiom.
standard math The commitment account of trust (Hawley 2014): to trust someone is to rely on them to fulfil a commitment.
Adopted in Section 5 from external literature (Hawley 2014) as the framework for analyzing trust in mathematics. Not defended within the paper.
domain assumption Formal verification via ITPs provides greater (though not absolute) confidence in correctness than informal community checking alone.
Treated as axiomatic throughout Sections 3-5. The paper acknowledges software bugs and translation gaps as residual risks but treats the overall reliability improvement as given.
ad hoc to paper The technological changes under discussion constitute a 'turn' rather than mere incremental adoption of tools.
This is the paper's central framing claim, stated in Section 1 and concluded in Section 7. It is not independently derived but asserted as an interpretive judgment.

invented entities (2)

Simil-proof independent evidence
purpose: An argument that conforms to the inferential standards of a legitimate mathematical community and is generally taken to be a proof by relevant experts, but which may contain subtle errors.
Introduced in prior work (De Toffoli 2021a) and applied here. It is a conceptual tool, not a physical entity. It has independent grounding in the historical record of erroneous proofs (Kempe's proof, cited in Section 3) and is falsifiable in principle: a simil-proof that is later shown correct becomes a proof, and one shown incorrect becomes an erroneous simil-proof.
Hybrid epistemic network
purpose: A conceptual model for mathematical collaboration where knowledge is distributed among human mathematicians, automated tools, and technological infrastructure, with no single node knowing all parts.
Introduced in Section 4 as a way to conceptualize projects like the Equational Theories Project. It is a descriptive framework rather than a measurable entity. Its usefulness depends on whether it generates predictions or insights beyond the 'social machine' concept of Martin and Pease (2013) that it extends.

how reviews work

0 comments

Cite this review

Pith. "Pith review of The Technological Turn in Mathematics." pith.science (2026). https://pith.science/paper/QCDQNEZP

@misc{pith2026260707162,
  author       = {Pith},
  title        = {Pith review of: The Technological Turn in Mathematics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QCDQNEZP}},
  note         = {Machine review of arXiv:2607.07162}
}

read the original abstract

Quickly evolving technologies, such as Interactive Theorem Provers (ITPs), Automated Theorem Provers (ATPs), and Large Language Models (LLMs), all falling under the general heading 'AI for mathematics,' are transforming mathematical practice in profound ways. This chapter explores the implications of these innovations, focusing on their impact on how mathematical knowledge is created and shared. It also discusses how they are reshaping the social dimension of mathematics, altering collaboration dynamics, trust relationships, and the collective production of knowledge. For instance, tools like ITPs facilitate large-scale collaborations and make new types of teamwork possible, where trust is not a necessary ingredient. ITPs also help us mitigate our human fallibility, yet they raise questions about the nature of formalization and the relationship between traditional and formal mathematics. Technologies such as LLMs are reshaping the division of epistemic labour between humans and machines and urge philosophers of mathematics to ask questions about the value of their work.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

3 extracted references · 3 canonical work pages

[1]

Deep Learning and Mathematical Intuition: A Review of (Davies et al. 2021)

"Argumentation Theory for Mathematical Argument. " Argumentation 33 (2): 173-214. Davis, Ernest. 2021. "Deep Learning and Mathematical Intuition: A Review of (Davies et al. 2021). " Preprint, ArXiv: 2112.04324v1. De Toffoli, Silvia. 2021a. "Groundwork for a fallibilist account of mathematics. " The Philosophical Quarterly 71 (4): 823-844. ---. 2021b. "Rec...

work page Pith review arXiv doi:10.1093/philmat/nkaf019 2021
[2]

Marton's Conjecture in abelian groups with bounded torsion

Springer Berlin Heidelberg. Gowers, Timothy, Ben Green, Freddie Manners, and Terence Tao. 2024. "Marton's conjecture in abelian groups with bounded torsion. " arXiv preprint arXiv:2404.02244. Gowers, Timothy, and Michael Nielsen. 2009. "Massively collaborative mathematics. " Nature 461 (7266): 879-881. Habgood-Coote, Joshua, Line Edslev Andersen, Rosemary...

work page Pith review arXiv 2024
[3]

Formal Mathematical Reasoning: A New Frontier in AI

"A Formal Proof of the Kepler Conjecture. " Forum of Mathematics, Pi 5. Hamami, Yacin. 2022. "Mathematical Rigor and Proof. " Review of Symbolic Logic 15 (2): 409- -449. Harris, Michael. 2015. Mathematics without apologies. Princeton University Press. ---. 2024. "Automation Compels Mathematicians to Reflect on Our Values. " BULLETIN (New Series) OF THE AM...

work page Pith review arXiv 2022

Pith tools

Reviewed July 9, 2026 · model on record in the stance chip above.

[1] [1]

Deep Learning and Mathematical Intuition: A Review of (Davies et al. 2021)

"Argumentation Theory for Mathematical Argument. " Argumentation 33 (2): 173-214. Davis, Ernest. 2021. "Deep Learning and Mathematical Intuition: A Review of (Davies et al. 2021). " Preprint, ArXiv: 2112.04324v1. De Toffoli, Silvia. 2021a. "Groundwork for a fallibilist account of mathematics. " The Philosophical Quarterly 71 (4): 823-844. ---. 2021b. "Rec...

work page Pith review arXiv doi:10.1093/philmat/nkaf019 2021

[2] [2]

Marton's Conjecture in abelian groups with bounded torsion

Springer Berlin Heidelberg. Gowers, Timothy, Ben Green, Freddie Manners, and Terence Tao. 2024. "Marton's conjecture in abelian groups with bounded torsion. " arXiv preprint arXiv:2404.02244. Gowers, Timothy, and Michael Nielsen. 2009. "Massively collaborative mathematics. " Nature 461 (7266): 879-881. Habgood-Coote, Joshua, Line Edslev Andersen, Rosemary...

work page Pith review arXiv 2024

[3] [3]

Formal Mathematical Reasoning: A New Frontier in AI

"A Formal Proof of the Kepler Conjecture. " Forum of Mathematics, Pi 5. Hamami, Yacin. 2022. "Mathematical Rigor and Proof. " Review of Symbolic Logic 15 (2): 409- -449. Harris, Michael. 2015. Mathematics without apologies. Princeton University Press. ---. 2024. "Automation Compels Mathematicians to Reflect on Our Values. " BULLETIN (New Series) OF THE AM...

work page Pith review arXiv 2022