AI-assisted maths on Zenodo, 17–19 Oct 2026
A claimed proof of the 1985 Arad–Herzog conjecture, Messerschmidt’s conjecture confirmed with Codex, and a Claude-run study on AI-written science
Awaiting review
The takeaway
Zenodo, not arXiv, carried the weekend’s AI-assisted results: an unknown author claims the 40-year-old Arad–Herzog conjecture on products of conjugacy classes with agentic OpenAI and Anthropic tools and a literature-relative Lean check, and a Bologna mathematician confirms Messerschmidt’s 2018 conjecture with Codex. None is refereed.
Status
- Claim
Awaiting review
- Our reporting
- Medium confidence
- Verification
- Lean-verified
- Importance
- 3 of 5
- Last verified
- 19 October 2026
Your AI and this story
- GPT-6 Astra172 days after its cutoff
- Claude Opus 5.5111 days after its cutoff
- Gemini 4 Argon202 days after its cutoff
- Grok 4.7141 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 111 days before it.
Key facts
- Arad–Herzog conjecture (Arad & Herzog, Products of Conjugacy Classes in Groups, LNM 1112, 1985): Aran S. Ziegler, ‘A proof of the Arad–Herzog conjecture’, Zenodo 23450488 (Oct 19, 2026; manuscript dated 18 Oct, ‘Private review version 3 … unreleased’; ~66 pp. plus a review-support zip). The paper finishes the remaining families on top of Fisman–Arad (1987, alternating groups), Moori–Tong-Viet (2011), Guralnick–Malle–Tiep (2013, projective linear and more) and Guralnick–Navarro (2016, square case). It also identifies an exceptional family of products with exactly two classes in odd-dimensional orthogonal groups over odd fields
- Arad–Herzog AI statement: ‘The author used agentic AI tools developed by OpenAI and Anthropic to carry out their research strategy and assist with Lean formalisation and manuscript preparation. The author takes full responsibility for the content.’
- Arad–Herzog Lean caveat (§12): the compiled declaration AH.FinalLiterature.global_literature_relative states the conjecture for all finite simple nonabelian groups, but its dependency closure contains 859 axioms. Besides the 3 standard Lean axioms, these are ‘587 binary ordinary imports and 269 other named mathematical inputs’ (published classification and character-theoretic results taken as axioms). The ‘revised verification artifacts are currently held in the private project archive’. The author is not otherwise known in the field, and there was no expert reaction as of Oct 19
- Messerschmidt’s 2018 conjecture confirmed: Martino Lupini (Università di Bologna, ERC grant DAT), ‘Lipschitz positive decompositions of von Neumann preduals’, Zenodo 23440415 (Oct 17). The selfadjoint predual of a von Neumann algebra has a Lipschitz positive decomposition iff the algebra is subhomogeneous, so infinite-dimensional selfadjoint trace class answers the Lipschitz decomposition problem negatively. The finite-dimensional constants satisfy c·log log N ≤ c_N ≤ C·log N. Disclosure: ‘OpenAI Codex was used in October 2026 … for mathematical exploration, proof development, source retrieval, manuscript drafting … and to assist in checking proofs’
- Lupini, ‘Common null complements for polynomial images of finite sums of perfect sets’, Zenodo 23440154 (Oct 17): for sums of D+2 compact perfect sets, one null F_σ set B gives P(C)+B = ℝ for every nonconstant polynomial of degree ≤ D, so Jung and Lai’s Conjecture 4.8 holds for such sums (it stays open in general). Codex was used ‘for mathematical exploration’ and ‘Separate Codex collaborators developed and reviewed arguments’
- Matching complexes: Zihuan Xu, ‘A Planar Counterexample to the Sphere-Wedge Conjecture for Matching Complexes’, Zenodo 23437631 (Oct 17). The matching complex of the planar pentagonal bipyramid collapses to a torus, which disproves the conjecture that matching complexes of planar graphs are contractible or wedges of spheres; it is the smallest possible counterexample, unique among 7-vertex planar graphs. ‘An OpenAI Codex assistant was used in developing the arguments, drafting the manuscript, and writing the verification programs’
- Cubelike Graph Core Conjecture (Mančinska–Pivotto–Roberson–Royle), partial: Qihang Wang, ‘Integral Cayley graphs and abelian rigidity in cores of cubelike graphs’, Zenodo 23445385 (Oct 18). It proves the conjecture for cores with a normal abelian Cayley representation and classifies the abelian groups on which every integral Cayley graph is cubelike. Models named: ‘OpenAI GPT-5.6 Sol, Anthropic Claude Fable 5, Anthropic Claude Fable 5.1, OpenAI GPT-6.1 Sol, OpenAI GPT-6 Astra, and Anthropic Claude Opus 5.5’; acknowledges Gewu Intelligence Lab (the same author and lab as arXiv 2610.20649 in the 16 Oct catalogue)
- Fluid dynamics: Adilkhan Salkimbayev, ‘Linear stability of asymmetric Burgers vortices at large circulation beyond the half-strain threshold’, Zenodo 23448800 (Oct 18). Maekawa’s vortex is linearly stable for every asymmetry λ ≤ 0.9399 at large circulation (previously known only for λ < 1/2), with interval-arithmetic certificates (python-flint). ‘The work was carried out with substantial help from generative AI (Anthropic’s Claude)’
Show 4 more
- Quantum chemistry: Jonas Matuzas, ‘Single-centre expansions of three-centre 1s Slater integrals: convergence rates and certified error bounds’, Zenodo 23448184 (Oct 18): ‘OpenAI ChatGPT-6 Pro was used for derivations, computations, and drafting; Claude Opus 5.5 (Anthropic) was used for independent re-derivation and audit’
- Minor: Lien-Hung Su, congruences for restricted-cycle and Pell-weighted permutations (OEIS A193437/A193438/A244432), Zenodo 23445724, ChatGPT used ‘for exploratory mathematical discussion’
- AI-run science as experiment: Diego Halabi, ‘Do cats rot your teeth?’, Zenodo 23458160 (v2 Oct 18, preregistration v1 doi 10.5281/zenodo.23435708, OSF RVH7J). It is a preregistered NHANES 2005–2006 study ‘produced almost entirely by an AI agent (Claude, model claude-opus-5-5, Anthropic, run in Claude Code)’ and released with the agent’s redacted session transcripts. The primary result was inconclusive (prevalence ratio 1.12, 95% CI 0.93–1.36). The preprint shows how 4 of 45 preregistered estimates ‘could support a positive paper without breaking the registration’ and argues that ‘verification, not writing, is now the bottleneck of scientific communication’
- Zenodo noise (sweep of 19 Oct): of ~50 Zenodo hits, 9 were serious records with AI statements (above), about 6 were serious-looking records with no AI statement found or not scanned (e.g. Weak Viterbo pentagonal certificate and (ℤ/4ℤ)^d sumsets: no statement in the PDF; substitutive points: not scanned), and the rest were crank physics/RH ‘proofs’, off-topic social-science surveys or matches with no AI involvement
What happened
The weekend of 17–19 Oct 2026 had no arXiv announcement, and the sweep’s new Zenodo step became the main source of AI-assisted results. Each candidate record’s PDF was downloaded and searched for an AI statement; the quotes above come from those PDFs or the record descriptions.
Arad–Herzog. In 1985 Zvi Arad and Marcel Herzog asked whether a product of two nonidentity conjugacy classes in a nonabelian finite simple group must contain at least two conjugacy classes. The question came out of Szép’s conjecture on products of centralisers. Alternating groups (Fisman–Arad 1987), sporadic groups and some low-rank cases (Moori–Tong-Viet 2011), projective linear groups and wide classical and exceptional ranges (Guralnick–Malle–Tiep 2013) and the square case (Guralnick–Navarro 2016) were already known. Ziegler’s ~66-page manuscript claims the remaining families and assembles the full statement over all finite simple groups. Readers should weigh three caveats: (1) the author has no visible track record in the field and the record calls itself a private, unreleased review version; (2) the Lean development proves the theorem relative to 856 named inputs, mostly published classification and character-table results imported as axioms, and its final verification artifacts are not public; (3) the author says agentic OpenAI and Anthropic tools carried out “their research strategy”, which suggests a largely agent-driven project. If experts confirm it, it is a notable AI-assisted resolution in finite group theory. Until then it is an unverified claim.
Codex as collaborator. Martino Lupini (Università di Bologna, holder of the ERC grant “Definable Algebraic Topology”) posted two Codex-assisted papers on Oct 17. One confirms Messerschmidt’s 2018 conjecture with a complete characterisation (subhomogeneous algebras) and near-optimal growth bounds. In the other, “separate Codex collaborators developed and reviewed arguments”. Zihuan Xu’s matching-complex counterexample follows the same single-author, Codex-assisted pattern. His same-day paper on substitutive points (Krawczyk’s Conjecture 3.4, Zenodo 23437635) was not scanned for an AI statement.
AI-run study. Diego Halabi’s “Do cats rot your teeth?” is a meta-experiment. A Claude Opus 5.5 agent in Claude Code ran a deliberately implausible epidemiology question from preregistration to preprint, and the record logs the time taken, human interventions, agent errors and the share of agent-written text. The point is that a by-the-book AI-written paper could be spun positive from 4 of 45 preregistered estimates.
Why it matters
Zenodo has become a channel for AI-assisted mathematics that skips arXiv moderation. That makes it the place to catch claims early, such as Ramanujan’s inequality on Oct 16–17 and Arad–Herzog now. Many of these claims carry heavy caveats. Literature-relative Lean proofs of the Arad–Herzog kind check the assembly of a proof, not the classification inputs.
What is disputed or not yet verified
| Verification | Unrefereed Zenodo records; Arad–Herzog has a literature-relative Lean development whose final artifacts are private; interval-arithmetic certificates for the Burgers result; no expert reactions yet |
|---|
Sources
10 sources from 2 sites. Numbers match the chips in the text.
10 sources: 10 primary
Primary
- Zenodo 23450488: A proof of the Arad–Herzog conjecture (Aran S. Ziegler, Oct 19)zenodo.org, paper
- Zenodo 23440415: Lipschitz positive decompositions of von Neumann preduals (Lupini)zenodo.org, paper
- Zenodo 23440154: Common null complements for polynomial images of finite sums of perfect sets (Lupini)zenodo.org, paper
- Zenodo 23437631: A planar counterexample to the sphere-wedge conjecture for matching complexes (Xu)zenodo.org, paper
- Zenodo 23445385: Integral Cayley graphs and abelian rigidity in cores of cubelike graphs (Wang)zenodo.org, paper
- Zenodo 23448800: Linear stability of asymmetric Burgers vortices beyond the half-strain threshold (Salkimbayev)zenodo.org, paper
- Zenodo 23448184: Single-centre expansions of three-centre 1s Slater integrals (Matuzas)zenodo.org, paper
- Zenodo 23445724: Congruences for restricted-cycle and Pell-weighted permutations (Su)zenodo.org, paper
- Zenodo 23458160: Do cats rot your teeth? Study produced by an AI agent (Halabi)zenodo.org, paper
- Guralnick, Malle & Tiep (2013): Products of conjugacy classes in finite and algebraic simple groups (arXiv 1202.2627)arxiv.org, paper
Changes
- Filed from the 19 Oct sweep (Zenodo step); PDFs of 13 candidate records scanned for AI statements