Will AI Kill the Research Paper? Mathematicians Offer a New Answer

The Pressure on the Paper System Was Already There—AI Just Accelerates It

On Lex Fridman's podcast, mathematicians raised a question that sounds almost absurd: Will AI kill the research paper? But the discussion quickly turned to a more grounded observation—the academic journal system was already under enormous strain long before AI entered the picture at scale.

According to views shared on the show, math journals have long faced a dual crisis of surging submissions and exhausted reviewers. Reviewers must expend significant effort to understand and verify theorems, and more and more people are producing knowledge far faster than the review system can digest it. One guest put it bluntly: "The only thing worse than this system is not having this system." Because mathematical epistemology—how we arrive at truth—inherently demands slowness, care, and precision.

And then he goes to the board and gives them a half hour lecture on how he does that.

This pressure isn't caused by AI, but AI-generated papers and proofs are making the review burden even worse. Someone on the show remarked: "Now we're faced with the task that someone's got to review all these AI-generated papers."

From Papers to Reusable Components: Lessons from Engineering

A key point of divergence in the discussion concerns the definition of "academic contribution." In engineering and applied sciences, the forms of contribution have long been diverse. Writing reusable code, maintaining a dataset, providing an API—any of these can be more valuable than a 15-page PDF. Especially in large collaborative projects—like LIGO or cancer research involving hundreds of people—the output has never been a single paper, but rather a collection of code, data, and reusable modules.

Guests argued that the AI era will make this trend even more pronounced: a small component, codebase, or dataset that can be directly invoked by the community or by AI may be worth more than a single result or a single paper. This raises a question worth pondering for pure mathematics: How should mathematical "reusability" be expressed?

Lean and Mathlib: The Formalization Revolution in Mathematics

The show highlighted Lean and Mathlib. Lean is an interactive theorem prover, originally designed for writing bug-free code. It works by checking proofs line by line—the computer determines whether each line follows from the previous line's goal state through logical inference. Mathlib is a large, general-purpose library for formalized mathematics within the Lean ecosystem, well-maintained and well-organized, so researchers don't have to start from Peano axioms to define natural numbers.

Guests noted that Lean and Mathlib are changing the logic of mathematical paper writing. Traditional math papers have a fundamental problem: there is no perfectly written theorem specification. Open any page of a math book to a lemma, and there's a massive amount of context that's been omitted—the author assumes you've read the previous 60 pages. This means mathematical results can't be used "out of the box" like physical components of a rocket.

But if mathematics becomes fully formalized in the future, specifications would be built into the objects themselves—Lean won't let you use a theorem incorrectly. This was Mathlib's dream ten years ago: to make mathematics run like science.

Interactive Theorem Prover (ITP) is a class of software tools in which users input mathematical definitions and proof steps, and the system checks in real time whether each step is logically valid. Unlike automated theorem provers, ITPs emphasize human-machine collaboration: humans provide proof strategies and key lemmas, while the machine handles verification details and prevents errors. Lean is a relatively new and active representative, relying on Dependent Type Theory as its foundation, which allows mathematical objects and propositions to be uniformly represented as types, with proofs being terms that construct those types. This "proofs as programs" correspondence (the Curry-Howard isomorphism) allows Lean to both verify mathematical proofs and be used for writing highly reliable software. Mathlib is a mathematical knowledge base built on top of Lean, currently containing extensive formalized content from basic algebra to cutting-edge research, organized like a software library with support for imports, reuse, and automated checking.

The GitHub Model and Open Source Community Fatigue

The discussion also extended to open source communities. The explosive growth of code on GitHub has overwhelmed maintainers—a flood of pull requests leaves a small number of core maintainers psychologically and physically drained. This fatigue closely mirrors the plight of academic reviewers: both depend on human motivation and enthusiasm, and once depleted, the system breaks down.

And now we're faced with the task that someone's got to review all these AI generated papers.

Mathlib's operating model was brought up as a contrast: not papers, but pull requests. Maintainers and reviewers offer suggestions in PRs, and the entire process is publicly visible—who contributed which idea, which simplification, where existing lemmas could be reused. This transparency is something traditional journal review cannot provide.

Pull Request (PR) is a collaboration mechanism on platforms like GitHub: contributors propose code changes, and maintainers review, discuss, and decide whether to merge. This model makes changes transparent, allowing anyone to comment and trace history. However, as open source projects scale up, PR volume surges, and core maintainers must invest substantial time in code review, responding to discussions, and resolving conflicts. Research shows that many critical open source projects are sustained by only a few volunteers, and long-term overload leads to maintainer burnout and even project stagnation. This closely parallels the dilemma of academic reviewers facing massive submission volumes: both rely on limited human attention and intrinsic motivation, lacking sustainable incentive and division-of-labor mechanisms.

AI Proofs Need New Verification Models

Regarding AI-generated mathematical proofs, guests held a cautious but open attitude. Natural language proofs cannot be trusted because these systems are stochastic processes—what they say doesn't mean they truly "understand" mathematics. But automated formalization has advanced to the point where AI can write Lean code that is checked layer by layer down to the axioms on top of Mathlib.

The show mentioned OpenAI's improvements to the fast Fourier transform and the phenomenon of people on Twitter接力 improving constants. A guest said they're not on Twitter and aren't quite sure how to regard these fragmented advances, but believes new models are needed to organize and consolidate these improvements rather than simply applying the old paper system.

Someone suggested that perhaps Mathlib's crowdsourcing model could be borrowed to let the community collaboratively verify AI proofs while preserving the rigor of traditional journals. This isn't about replacing the old system, but about relieving pressure it can no longer bear.

me about lemma 57,

Papers Won't Disappear, But Their Form Will Change

Synthesizing the views of multiple mathematicians, papers as a knowledge carrier won't suddenly die, but they are no longer the only—or even necessarily the most important—form of contribution. Formalization tools allow technical details to be machine-verified, freeing researchers to focus more on communicating and explaining ideas.

One guest mentioned that their favorite reason for giving research talks is the ability to strip away the messy complexity of a paper and highlight the truly exciting core idea and special cases. When Lean takes over technical verification, writing itself may return to something more essential: clearly explaining why this idea matters.

I don't know,

Mathematics' culture of independence is also a variable. Mathematicians generally dislike being directed and prefer solitary work, which creates tension with the collective organizational models required by large collaborative projects. How to establish new collaborative verification mechanisms while preserving individual creativity is a question the mathematical community must answer in the AI era.

One-line Summary

AI intensifies pressure on the academic paper system; mathematics needs to shift toward formal verification and diverse forms of contribution.

Paragraph Summary

Based on discussions among mathematicians on the Lex Fridman podcast, the article notes that the academic journal system was already overwhelmed by surging submissions and reviewer fatigue before AI entered the scene, and AI-generated papers are making things worse. Engineering has long treated code, datasets, and APIs as important contributions, and this trend may extend to mathematics. Lean and Mathlib provide built-in specifications for mathematical theorems through formal verification, avoiding the context-loss problem of traditional papers. The PR overload in open source GitHub communities structurally resembles the academic review dilemma, calling for new collaborative verification models. AI-generated proofs can't rely on natural language, but can be verified layer by layer through automated formalization into Lean code. Papers won't disappear, but their form will change, allowing researchers to focus more on communicating ideas.