AI Mathematical Proofs Explained: What Lean Can Verify and Why Experts Still Matter

Research release: OpenAI's public mathematics repository currently lists 719 manuscripts organised into 372 families of related results. Its revision history records the withdrawal of three manuscripts after an error affected an argument and dependent papers. These are research claims at different stages of checking, not a declaration that every open problem mentioned has been conclusively solved.

Verification status: The repository reports 300 of 719 top-line results as formalised in Lean, approximately 42%. A family may contain a principal result, supporting arguments or alternative proofs; manuscripts, families and independently verified discoveries are therefore different counts.

Conjecture: A mathematical proposition believed to be true but not yet established by an accepted proof.

Proof: A logical argument deriving a conclusion from clearly stated assumptions, definitions and previously established results. Checking many numerical examples can suggest a pattern, but cannot by itself prove a claim about infinitely many cases.

Theorem: A proposition established through proof within a stated mathematical framework. Its assumptions matter: a result proved for positive integers cannot automatically be asserted for all real numbers.

Illustration: For integers m and n, two even numbers can be written as 2m and 2n. Their sum is 2(m+n), which is even because m+n is an integer. This explains the general result rather than merely listing examples.

Plausible language: A language model generates sequences using learned patterns. Fluent mathematical prose can conceal an unjustified inference, an incorrect citation or a missing assumption.

Dependency risk: A long argument may use several earlier results. If one supporting result fails, multiple dependent claims may require revision.

Different tasks: Finding an answer, producing a readable argument, checking each inference and assessing a result's importance are distinct activities. Strong performance on one does not automatically establish the others.

Proof assistant: Lean is an open-source programming language and interactive theorem prover. It represents definitions, statements and proofs in a precise formal language.

Kernel: A small trusted checking component checks whether the submitted proof follows the formal rules. Automation can help construct a complicated proof, while the checking stage validates the resulting formal object.

Human participation: Researchers choose statements, define objects, guide difficult steps and interpret results. A proof assistant supports these tasks; it does not by itself decide which research question is valuable.

Formalisation: Translating a mathematical statement and argument into a language that a proof assistant can check.

Autoformalisation: Using AI or other automation to assist that translation. This can reduce the considerable effort required to convert informal papers into complete formal developments.

Translation problem: If an informal claim concerns every case but its formal version covers only a restricted case, the formal proof may be valid while the advertised claim remains unproved. The statement being verified must therefore be compared with the original claim.

Logical guarantee: A properly checked proof establishes the encoded conclusion under the encoded assumptions and accepted foundational axioms. This is stronger than asking another chatbot whether an argument seems correct.

Scope limit: It does not automatically establish that the formal statement faithfully represents the intended problem, that the result is new, or that its explanation is useful.

Independent scrutiny: A recent research preprint by Alexander Bastounis, Fabian Circelli and Anders C. Hansen examines mismatches between natural-language arguments and their AI-generated Lean translations. Its central warning concerns translation fidelity; it should not be interpreted as showing that all proof assistants are unreliable.

Reproducibility: Reviewers should obtain the complete project, with specified software and library versions, and independently rebuild it.

Actual coverage: The build must check the file containing the claimed theorem. Merely placing Lean code beside a manuscript does not establish verification.

Assumption audit: Lean permits additional axioms and unfinished proof placeholders such as “sorry”. Reviewers must inspect dependencies and ensure the conclusion has not simply been assumed.

Semantic review: A specialist must check definitions, quantifiers and hypotheses against the intended theorem. Reproducibility, logical checking and faithful representation answer different questions.

Shared foundation: Mathlib is a community-maintained Lean library of formal mathematical definitions and results. Researchers can reuse checked components rather than repeatedly rebuilding elementary foundations.

Knowledge infrastructure: A reliable library makes collaboration easier, but additions still require review, documentation and maintenance. Its value comes from reusable verified knowledge rather than the number of uploaded files alone.

Correctness versus contribution: A valid proof may establish an already known result, an unimportant special case or a statement made trivial by its assumptions. Researchers assess novelty, significance and connections with existing knowledge.

Credit: AI-assisted work should acknowledge relevant literature, human contributions and the role of automated tools. Otherwise priority disputes and inaccurate claims of discovery can distort incentives.

Research governance: The independent Advisory Group on Mathematics and Artificial Intelligence stresses responsible release and community-led understanding. Its advisory involvement is not certification of the correctness or impact of a company's results.

Scientific capability: AI-assisted reasoning and verified software can support research in mathematics, computing and engineering. Potential applications include checking algorithms, cryptographic protocols and safety-critical systems, although real deployments require domain-specific validation.

Access: If advanced models, computing resources and verification expertise are concentrated in a few firms, smaller universities may struggle to participate. Shared infrastructure and training can help reduce that gap.

Policy link: The IndiaAI Mission includes compute capacity, skills and Safe & Trusted AI among its pillars. Research support should combine access to tools with training in logic, reproducible methods and independent evaluation.