·pro
Book a call
Back to Writing The Most Important Lemma of 2026 Was Written by an AI. There's No Way to Cite It.

The Most Important Lemma of 2026 Was Written by an AI. There's No Way to Cite It.

9 days ago 14
AI & LLMsData & MathematicsStrategy & Career

May 2026. Four number theorists upload a paper to arXiv.

Thomas Bloom, Will Sawin, Carl Schildkraut, and Dmitrii Zhelezov have just disproved the Erdős–Szemerédi sum-product conjecture over the real numbers — a problem that had stood unsolved for over forty years. The proof is clean, elegant, surprising. The kind of result that makes mathematicians stop what they're doing and read.

But buried in Section 1.3, under a modest subheading titled "The role of AI in this proof," is something stranger still:

GPT-5.5 Pro was used as a sounding board in the early stages of the development of this proof, but the final proof, including all the main ideas, was almost entirely human-generated (the exception being the suggestion of Lemma 3.4, which replaced a more complicated result of Schinzel with a short elementary argument).

Lemma 3.4. The golden-ratio separation lemma. A four-line argument using the golden ratio to partition real numbers into two sets. It replaced pages of heavy Schinzel machinery that would have added significant complexity to the proof. Without it, the result might still have happened — but it would have been messier, harder to read, and harder to verify.

That lemma was suggested by a large language model.

And now it is un-citable.

Contents

arXiv has no metadata field for AI contributions. The DOI system has no entity type for non-human authors. OpenAlex doesn't index acknowledgments. Semantic Scholar can't parse "GPT-5.5 suggested this" as a citation edge. If you want to build on Lemma 3.4, you reference "Bloom et al. (2026)" — the paper, not the proof. The entity that discovered the lemma disappears from the graph entirely.

This is not a philosophical problem about whether AI can be creative. It's an infrastructure problem. Our citation systems — designed in an era when all intellectual contributions came from named humans — have no mechanism to track, attribute, or amplify non-human contributions.

And it's already breaking the literature.

The Four-Line Argument That Changed Everything

To understand why this matters, you need to understand what Lemma 3.4 actually did.

The Erdős–Szemerédi sum-product conjecture, proposed in 1983, asks whether you can construct a set of real numbers that's simultaneously well-behaved under both addition and multiplication. The conjecture says no: any finite set of real numbers must have either a large sumset or a large product set — you can't have both small. For four decades, mathematicians chipped away at the exponent, getting closer and closer to the conjectured bound.

The Bloom–Sawin–Schildkraut–Zhelezov proof broke the conjecture entirely. The key technical move was a separation argument: if you can split your number set into two pieces in a clever way, you can analyze each piece independently and show that at least one of them behaves badly. That's where Lemma 3.4 comes in.

The original approach used a result due to Schinzel — a deep, nontrivial theorem from analytic number theory requiring substantial machinery to invoke. The AI-suggested replacement used only the golden ratio and elementary arithmetic. It's the kind of argument a graduate student could verify in five minutes.

Tim Gowers, the Fields Medalist, was one of the first people to read the paper. He noted that the AI contribution wasn't peripheral. It was load-bearing.

That's the paradox at the center of this story: the most cited beneficiary of Lemma 3.4 will be "Bloom et al. (2026)." The actual originator won't appear in a single citation record.

How We Got Here: A Short History of an Accelerating Trend

The sum-product paper wasn't an anomaly. It was an endpoint — for now — of a trend that began quietly in late 2024.

2024: The Opening Move

Robert Ghrist, a topologist at Penn, was working on lattice-valued network flows with collaborators Julian Gould and Miguel Lopez. They had a human proof. Out of curiosity, Ghrist fed GPT-o1-mini a failed proof attempt from Claude and asked it to diagnose the error.

The model spent 43 seconds "thinking" — the internal chain-of-thought that o1-series models perform before outputting a response — and produced an entirely new proof. The authors called it more beautiful than their own.

The resulting paper (arXiv:2410.00315) included a detailed appendix documenting the AI collaboration: the prompts, the failures, the breakthrough moment. At the time, it read like a novelty. An interesting case study. A footnote.

In retrospect, it was the opening move of a very fast game.

2025: The Tipping Point

In August 2025, a group of researchers published a paper challenging LLMs with a group theory problem: prove that the presentation ⟨x, y | xy² = y³x, yx² = x³y⟩ defines the trivial group. Their conclusion was blunt: "No LLM could solve it."

Two months later, Bartosz Naskręcki handed GPT-5 Pro the same problem. The model produced a complete proof in 14 minutes, using a purely algebraic approach no human had considered. The gap between "impossible" and "routine" was exactly two months of model generations.

Around the same time, Tim Gowers was working on a proof when he hit a gap. He knew what lemma he needed but didn't know how to prove it. He queried GPT-5. The model returned a correct proof in 20 seconds, relying on a lemma Gowers had never encountered. His assessment, posted on X:

"We are entering the brief but enjoyable era where our research is greatly sped up by AI but AI still needs us."

Gowers' phrasing was careful: brief but enjoyable. He saw the trajectory. He just didn't know how brief.

Spring 2026: The Breakout Quarter

Three events in rapid succession changed the landscape permanently.

The Unit Distance Conjecture. OpenAI's o1-series — a general-purpose reasoning model, not a specialized mathematical system — autonomously produced a complete proof of an 80-year-old problem in discrete geometry (the Erdős unit distance problem). The proof connected class field towers, Hilbert class fields, and Poitou–Tate duality: an unexpected bridge between discrete geometry and algebraic number theory that no human had previously seen.

The result was verified by nine leading mathematicians, including Alon, Bloom, Gowers, Litt, Sawin, Shankar, Tsimerman, Wang, and Wood. Gowers said he would recommend it for publication in the Annals of Mathematics — the most prestigious mathematics journal in the world.

Will Sawin simplified the proof and made it explicit in arXiv:2605.20579.

Erdős Problem #1196. GPT-5.4 Pro, given a single prompt and roughly 80 minutes of chain-of-thought reasoning, suggested a novel proof method combining von Mangoldt weights with Markov chains in the context of primitive sets. Terence Tao — arguably the greatest living mathematician — noted that human researchers "made a slight wrong turn at move one" on this problem, following a path that seemed natural but was subtly wrong. The AI arrived at the problem fresh, without those wrong-turn intuitions, and found the correct route.

What happened next is unprecedented: the AI-generated proof method was adapted to solve a related 60-year-old problem, the Erdős–Sárközy–Szemerédi conjecture. Lichtman, Tao, and collaborators formalized both results, including a Lean proof of the main theorem.

This is, to my knowledge, the first documented case of an AI-generated proof method having cascading downstream impact — a single AI insight solving multiple open problems.

The Sum-Product Conjecture. The story that opened this article. May 2026. GPT-5.5's Lemma 3.4 became the first non-human lemma formally credited in a major mathematical result. The breakthrough caught the attention of the wider scientific community, featuring prominently in Nature's coverage of how AI is transforming mathematical research (Castelvecchi, 2026).

Where This Ends: Feng26

In January 2026, DeepMind's Aletheia system — built on Gemini Deep Think — produced a complete research paper in arithmetic geometry. Not a proof assistant. Not a lemma suggestion. A complete paper: from raw problem data through iterative generation, verification, and revision loops, grounded in Python code and live search.

The paper was uploaded to arXiv as 2601.23245. The sole listed author is "Feng" — the human advisor who contextualized and submitted the work.

The mathematics is correct. The paper is publishable. And there is no mathematical author in any meaningful sense.

Five Levels of AI Contribution

Before we can fix the attribution problem, we need language for what we're tracking. Across roughly a dozen documented papers from 2024 to 2026, a clear taxonomy emerges — not a capability ladder, but a map of different modes of collaboration, each with its own attribution challenge.

Level What the AI Does Canonical Example
L1: Sounding Board Brainstorms directions; human filters and verifies Sum-product conjecture general development
L2: Lemma / Method Generation Produces a specific, verifiable argument Lemma 3.4; Gowers' GPT-5 lemma
L3: Full Proof Generation Generates a complete, independently verified proof Unit distance conjecture; Tsumura's 554th
L4: Human–AI Collaboration Iterative symbiosis; AI explores, human steers Erdős #1196 (Lichtman–Tao–GPT)
L5: Fully Autonomous Research AI produces a paper with zero human mathematical input Feng26 (Aletheia)

L1 and L2 are where the citation problem bites hardest today. A specific, verifiable, load-bearing piece of mathematics with no author in the author list. L3 and L4 push the question further: should an AI-generated proof be cited as the proof, or as the AI's output? L5 takes it to the limit: if no human contributed mathematically, who is the intellectual ancestor of the result?

Currently, the answer to all of these is the same: the human authors of the paper. Which is wrong.

The Infrastructure Gap

Every modern citation system — CrossRef, arXiv, OpenAlex, Semantic Scholar, Web of Science — was built on a single foundational assumption: every citable entity has at least one human author. This assumption is so deeply embedded that most of these systems don't even expose the concept of "contributor type" as a field.

That produces three structural failures:

No entity type for non-human contributors. An article has authors. A dataset has creators. A software package has maintainers. An AI-generated lemma has nothing — no entity type, no metadata field, no way to represent "GPT-5.5 Pro suggested this argument" as a structured fact in the citation graph.

No versioning for ephemeral models. GPT-5.5 Pro and GPT-5.4 Pro are different models with different capabilities. An o1-series proof differs in character from an o3-series proof. When a paper cites "OpenAI o1," it points to a snapshot that cannot be reproduced, audited, or updated. Human authors stay fixed; model versions drift, are updated, and eventually deprecate.

No mechanism for downstream impact tracking. The Erdős #1196 → Erdős–Sárközy–Szemerédi chain is a clear case of intellectual genealogy: one AI insight produced two solved conjectures. But OpenAlex won't draw that edge. Semantic Scholar won't surface it. The citation graph will show "Lichtman et al. (2026)" with no indication that a non-human entity was the intellectual origin of the proof method.

If this sounds familiar, it should. This is structurally identical to the alpha attribution problem in multi-strategy quantitative trading: if three models contribute signals to a portfolio and the risk system attributes the returns to the wrong one, you can't improve what you can't measure. Citation infrastructure for AI contributions is exactly the same class of problem as model attribution in finance. The asset is intellectual, not financial — but the tracking requirement is identical.

Of the twelve papers in the documented dataset of AI-credited mathematical contributions, exactly zero have structured metadata capturing those contributions. The signal lives in acknowledgments sections and X threads. Readable by humans. Invisible to machines.

What Structured Attribution Could Look Like

The solution doesn't require inventing a new standard. It requires extending existing ones.

A lightweight metadata schema for AI contributions could look like this:

{
  "ai_contributions": [
    {
      "model": "GPT-5.5 Pro",
      "version": "2026-05-15",
      "provider": "OpenAI",
      "contribution_type": "lemma_suggestion",
      "contribution_detail": "Suggested Lemma 3.4 (golden-ratio separation argument)",
      "credit_location": "Section 1.3"
    }
  ]
}

arXiv could add this as a metadata field. OpenAlex could index it as a work attribute. Semantic Scholar could create a "non-human contributor" entity type — a directed edge in the knowledge graph labeled "contributed_to" rather than "authored." CrossRef already supports contributor roles; adding "ai-contributor" is a schema change, not an architectural one.

The technical work is small. The hurdle is institutional: getting publishers, preprint servers, and indexing services to agree that non-human contributions exist, matter, and need to be tracked.

That conversation is starting. It's moving slowly. Meanwhile, the literature is accumulating edges that the graph can't see.

The Deeper Question

The unit distance proof revealed something beyond mathematics. The AI discovered a connection between discrete geometry and class field theory that no human had made. These fields were considered separate. The model didn't know that. It saw a path across a gap that human expertise had learned not to look across.

When mathematicians acquire deep expertise, they also acquire deep intuitions about what shouldn't work — about which approaches are "in the wrong direction." Those intuitions are usually correct. Occasionally they're wrong in exactly the way that makes a 50-year-old problem seem harder than it is.

An AI without those intuitions — without the accumulated sense of what looks "wrong" — explores differently. It doesn't make the right turns because it doesn't know which turns are canonical. Sometimes that blindness is a liability. Sometimes it finds a path through a wall that experts had learned to walk around.

Gowers called this era "brief but enjoyable." As of June 2026, we have:

  • Twelve or more documented papers crediting AI for substantive intellectual contributions
  • One fully autonomous AI-generated paper with a human figurehead author
  • Two solved 60-year-old conjectures from a single AI-suggested proof method
  • One AI-generated lemma that became essential to disproving a half-century conjecture

The brief era may already be ending. What comes next isn't AI that assists mathematical research. It's AI that is mathematical research, with humans in the role of editors, verifiers, and curators.

And our infrastructure — built for a world where all intellectual contributions came from named humans with ORCID IDs — is going to have to catch up.

The Lemma That Vanishes

Lemma 3.4 is four lines. It uses the golden ratio — a number known since antiquity, famous for appearing in art, architecture, and plant growth patterns — to make a clean separation argument in additive combinatorics. It's elegant in the way that good mathematics is: you see it and think of course, and then you wonder why no one found it before.

Other mathematicians will use it. They will read Bloom et al. (2026) and learn that separation argument and apply it to their own problems. When they do, they'll cite "Bloom et al. (2026)" — the paper. The actual originator of the argument will earn no citation index bump, no DOI, no way to measure its intellectual descendants. The database will record a human-to-human edge that obscures the actual provenance.

The AI wrote the lemma. The humans verified it. But the machine that discovered it won't appear in a single citation graph for the rest of the decade.

That's not a bug in AI. It's a bug in how we count what counts.

And unlike most infrastructure bugs — the kind that stay invisible until a system breaks — this one is already visible. The literature is already filling with AI contributions that the citation graph treats as human. Every passing month makes the historical record a little more wrong.

The question isn't whether to fix this. The question is how fast we decide that it matters.

Get in Touch

Interested in how AI is reshaping research, mathematics, and the infrastructure of knowledge? Want to discuss AI strategy, capability evaluation, or the broader implications of AI in technical domains?

Connect with me:

Whether you're thinking about AI's role in research, building AI-augmented workflows, or just want to discuss how the future of knowledge attribution is going to unfold — I'd love to hear from you.

References: Bloom et al. — arXiv:2605.28781 · Sawin — arXiv:2605.20579 · Alon, Bloom, Gowers et al. — arXiv:2605.20695 · Ghrist, Gould, Lopez — arXiv:2410.00315 · Feng (Aletheia) — arXiv:2601.23245 · Castelvecchi — Nature News: "'It is incredible': How AI is transforming mathematics" · Gowers (@wtgowers), Naskręcki (@nasqret), Lichtman (@jdlichtman), Sawhney (@mehtaab_sawhney), Bubeck (@SebastienBubeck) — X/Twitter, 2025–2026.