Lech Mazur’s AI-Assisted Sendov Conjecture Proof Changes What Counts as Evidence

Lech Mazur has announced an AI-assisted, Lean-verified proof of the sendov conjecture, a 67-year-old problem that resisted a complete argument until August 2026. The claim arrived with an unusual conflict built into it. A machine-checked proof offers stronger logical assurance than an ordinary draft, yet mathematicians still need to inspect what the machine actually checked.

Terence Tao then published a detailed mathematical digestion of the argument. That matters because Tao had already proved the conjecture for all sufficiently large polynomial degrees in 2020. His new exposition does not make him the original solver or remove every review question. It does show that a leading expert found a coherent mathematical mechanism worth reconstructing for human readers.

The real story is therefore larger than another difficult problem falling to AI. Mazur’s result tests a new division of labor among conjecture selection, AI-guided search, formal verification, and expert explanation. If the full proof survives continued scrutiny, those four stages will matter more than the simple label “AI solved it.”

What Changed in the Sendov Conjecture

The new claim closes the finite-degree gap left after decades of partial results, while attaching a machine-checkable certificate to the proposed proof.

The sendov conjecture concerns the relationship between a polynomial’s zeros and its critical points. A critical point is a zero of the derivative, so it marks where the polynomial’s local behavior changes.

Suppose every zero of a complex polynomial lies inside or on the unit disk. The conjecture says that each zero must have a critical point within distance one. The statement is elementary enough to draw, but a general proof remained elusive.

Blagovest Sendov proposed the problem in 1958, according to the historical account in Tao’s high-degree paper. Early literature sometimes attributed it to Lubomir Ilieff, which explains the older name Ilieff-Sendov conjecture.

Researchers gradually established the claim in restricted settings. The conjecture was known for degrees below nine, for special zero locations, and for several degree-dependent regions. Those results covered important territory without connecting every case.

Tao changed the landscape in December 2020. He proved that an absolute threshold exists beyond which every polynomial satisfies the conjecture. The result appeared in the 2022 volume of Acta Mathematica.

That theorem settled all sufficiently high degrees, but it did not supply a practical numerical threshold. Its compactness arguments established existence without producing a manageable cutoff. The full conjecture therefore did not follow by checking a clearly bounded list of remaining degrees.

Mazur’s August 2026 announcement claims to remove that gap with an argument formalized in Lean. Lean is a proof assistant that reduces a proof to definitions and logical steps checked by a small verification kernel.

That distinction matters. A conventional manuscript asks referees to follow prose, fill minor omissions, and verify calculations. A Lean proof asks software to reject any step that does not follow from the encoded assumptions and previously accepted results.

Machine checking does not turn a theorem into an unquestionable fact. It does substantially change the first verification question. Critics must identify an error in the formal statement, its definitions, its trusted dependencies, or the connection between the formal theorem and Sendov’s original claim.

Tao’s subsequent digestion adds a second form of evidence. His Sendov analysis reconstructs the formal result as recognizable mathematics and examines the proof’s central ideas.

Tao describes the proof as remarkably elementary. According to his account, it needs no substantial complex analysis beyond the fundamental theorem of algebra and basic facts about Möbius transformations.

The deepest named inequality is the Maclaurin inequality, which compares symmetric means of nonnegative numbers. That is unexpected because previous progress used sophisticated analytic, geometric, and asymptotic methods.

The event is best dated to August 2026, not to Tao’s 2020 result. Tao established the high-degree theorem in December 2020. Mazur announced the claimed complete, AI-assisted formal proof in August 2026, followed by Tao’s public digestion.

Why a Simple Statement Survived for 67 Years

Sendov’s problem remained open because local geometry around one zero must be controlled using information distributed across every zero and critical point.

The conjecture sounds like a nearest-neighbor claim. Pick one zero, draw a radius-one disk, and find a critical point inside it. However, a polynomial’s derivative depends on the entire configuration of zeros.

The Gauss-Lucas theorem provides the broadest geometric constraint. It says every critical point lies inside the convex hull of the polynomial’s zeros. When all zeros occupy the unit disk, all critical points remain there too.

That is not enough for Sendov’s claim. A critical point can lie inside the global convex hull while remaining more than one unit from a particular zero. Sendov demands a separate local guarantee for every zero.

The difficulty becomes clearest near the boundary. A chosen zero can sit close to the unit circle, while most critical points cluster elsewhere. A proof must rule out configurations that almost violate the desired distance.

Tao’s 2020 work explains why those near-counterexamples matter. His analysis divided the high-degree problem according to the selected zero’s position, then used different tools near the origin and boundary.

For zeros near the boundary, Tao refined perturbative arguments developed by earlier researchers. Near the origin, he used compactness, balayage, and the argument principle. Balayage is a method for replacing a distribution with boundary data while preserving its external potential.

Those methods proved that counterexamples cannot persist when the degree becomes large. Yet they did not produce an explicit threshold suitable for finishing the remaining cases by computation.

The new proof reportedly takes another route. Tao’s reconstruction reframes a supposed counterexample and extracts algebraic inequalities that its zeros and critical points must satisfy. The contradiction then emerges through elementary transformations and symmetric inequalities.

This mechanism matters more than the age of the problem. AI systems often perform best when they can search many algebraic reformulations, test intermediate lemmas, and receive exact feedback from a verifier.

A human mathematician can also explore those branches. The difference is the scale and speed of iteration. A formal agent can propose a step, compile it, study the failure, and try another formulation repeatedly.

That process fits Sendov’s problem unusually well. The statement is compact, many equivalent normalizations exist, and the goal can be expressed precisely. Each candidate inequality gives a verifier a clear pass-or-fail obligation.

The proof’s elementary character should not be confused with an easy discovery. Many famous arguments look simple after the right representation has been found. The difficult work often lies in locating the representation that makes the contradiction visible.

This is also why “AI searched harder” is an incomplete explanation. Search becomes useful only when the system has a productive formal language, a tractable target, and feedback capable of rejecting false moves.

Lean provides that feedback after formalization. Mazur provides problem selection, direction, interpretation, and accountability for the claim. Tao’s exposition supplies a human-readable route through the resulting artifact.

Those roles do not collapse into one authorless machine event. They form a pipeline, and each stage addresses a different source of uncertainty.

AI Generation Versus Formal Verification Is the Real Contest

The primary conflict is not AI against mathematicians, but generated reasoning against evidence that independent systems and experts can audit.

A language model can produce a polished proof containing a fatal gap. Mathematical prose is especially vulnerable because a false transition can resemble thousands of valid arguments in its training data.

Asking another language model to review the same proof does not fully solve the problem. Models can share training sources, reasoning habits, and blind spots. Their agreement may reflect correlated error rather than independent confirmation.

Formal verification changes the structure of that evaluation. Lean does not accept an argument because it sounds familiar. Its kernel checks whether each term has the required type under the stated definitions and axioms.

That gives the formal proof a harder evidentiary foundation than an unaudited chat transcript. It does not mean Lean understands mathematical importance, historical priority, or whether the chosen formal statement matches researchers’ intentions.

This boundary is essential. A proof assistant can verify the wrong theorem perfectly. A subtle mistranslation might weaken a hypothesis, alter a distance convention, or restrict the class of polynomials without making the formal file fail.

Formalization therefore creates two verification layers. The first asks whether the Lean code compiles under its trusted environment. The second asks whether the encoded theorem faithfully represents the sendov conjecture.

The second layer still needs mathematicians. Experts must inspect definitions, theorem statements, imported results, and any assumptions hidden behind abstractions. They must also compare the artifact with the conventional formulation.

Tao’s digestion is important at precisely this boundary. He translates the proof back into ordinary mathematics, identifies its mechanism, and relates it to the established literature.

That is different from lending celebrity approval to a headline. A mathematical digestion exposes structure that other specialists can challenge. It lets readers ask where each inequality enters and whether any case has disappeared during translation.

Mazur’s public role also matters. The phrase “AI-assisted” covers a wide range of workflows, from brainstorming to autonomous formal search. A responsible account should identify which steps came from AI, which came from humans, and which were checked mechanically.

The current evidence supports a cautious formulation. Mazur announced a complete proof, the associated artifact was presented as Lean-verified, and Tao produced a serious mathematical exposition. Those facts justify attention without making peer review irrelevant.

The strongest claim is not that an AI independently woke up and solved a famous conjecture. The stronger, better-supported conclusion is that an AI-enabled workflow produced a formal result that a top specialist could meaningfully digest.

That is already a significant change. Earlier AI math demonstrations often depended on benchmark problems with known answers or carefully prepared formal statements. Sendov was a recognizable open conjecture with extensive specialist literature.

Recent mathematical AI projects show the same verification-centered pattern. Aristotle, developed by Harmonic, has been used to search for and formalize proofs in Lean. A January 2026 Erdős resolution credited GPT-5.2 Pro, Aristotle, and human operator Kevin Barreto as separate contributors.

The Sendov case extends that model to a more prominent analysis problem. It also makes the collaboration chain unusually visible: conjecture, operator, AI search, proof assistant, and expert exposition.

This distribution of credit will become contentious. Mathematical authorship traditionally combines idea generation, proof construction, error checking, exposition, and historical positioning. AI-assisted formal work can distribute those functions across different people and systems.

Readers should resist two equally weak narratives. One treats the result as worthless because AI participated. The other treats formal compilation as proof that human mathematical judgment no longer matters.

The evidence supports a narrower conclusion. Generated proofs become much more credible when they meet a verifier, but their meaning still depends on faithful specification and expert interpretation.

What the Lean Certificate Does Not Settle

A verified artifact can establish logical validity while leaving specification, provenance, novelty, and scholarly acceptance open for review.

The first uncertainty concerns the exact theorem statement. Independent Lean users should compile the artifact, inspect its assumptions, and confirm that its definitions match the standard closed-unit-disk formulation.

This is not a procedural technicality. Formal proofs derive their force from exactness. A one-character change in an inequality can separate Sendov’s full claim from a nearby statement that was already known.

The second uncertainty concerns dependencies. Lean proofs commonly import established libraries containing algebra, topology, analysis, and finite constructions. Reviewers should identify any custom axioms, placeholders, or unproved declarations.

A clean kernel check is strong evidence only within the trusted computing base. That base includes Lean’s kernel, the formal source, and the hardware and software executing it. It is small compared with ordinary mathematical trust, but not nonexistent.

The third issue is provenance. “AI-assisted” should describe the workflow rather than serve as a promotional category. Researchers need enough detail to understand whether AI found the central idea, filled formal gaps, translated prose, or explored alternatives.

That information affects scientific interpretation. A system that autonomously finds a decisive lemma demonstrates a different capability from a system that formalizes a human-written argument.

Both uses remain valuable. They simply answer different questions about AI’s research capacity.

The fourth issue is novelty. AI systems can rediscover forgotten results or reproduce ideas embedded in obscure literature. Tao has repeatedly emphasized the importance of literature search when assessing machine-generated mathematics.

Sendov’s conjecture has accumulated decades of partial proofs, claimed proofs, and technical variants. Specialists must compare Mazur’s route with prior work before assigning historical credit to every component.

The fifth issue is exposition. A formal proof can be correct yet difficult to understand. Mathematics advances through reusable concepts, not merely certificates that a statement follows from axioms.

Tao’s digestion addresses this problem by compressing the formal chain into a human argument. Other mathematicians now need to test whether that account can be simplified, generalized, and taught without depending on the original search process.

The sixth issue is conventional peer review. A journal referee does more than check logical validity. The referee evaluates originality, clarity, citations, scope, and the relationship between claims and evidence.

A public expert reconstruction can accelerate that process, but it does not replace it. Neither social-media enthusiasm nor skepticism should be mistaken for a completed scholarly assessment.

The strongest skeptical position is therefore not “the proof is probably false.” Available evidence is more substantial than a typical online proof claim. The responsible skepticism concerns correspondence and completeness around the formal artifact.

Does the Lean theorem exactly encode Sendov? Does the file compile independently? Are all imports and assumptions acceptable? Does the informal explanation cover the same scope?

Those are answerable questions. That is an improvement over disputes about lengthy prose proofs, where disagreements can persist around implicit steps and competing interpretations.

A formal artifact gives critics a precise target. If an error exists, they can identify a definition, assumption, import, or translation. If repeated audits find none, confidence should rise accordingly.

Terence Tao’s Role Is Validation, Not Co-Ownership

Tao supplied crucial expert interpretation, but the public record distinguishes his earlier partial theorem from Mazur’s claimed complete proof.

Headlines that say “AI, Lech Mazur, and Terence Tao solved Sendov together” blur three separate contributions. That framing is understandable, yet mathematically inaccurate.

Tao’s 2020 theorem established the conjecture for sufficiently large degrees. It was a major partial result and transformed the remaining problem into a finite question in principle.

However, Tao’s proof did not resolve every degree. Its threshold was existential rather than explicit, so researchers could not simply enumerate the remaining cases.

Mazur’s announced proof targets the entire conjecture. AI-assisted search and Lean verification are central to that new claim. Tao entered afterward as an expert reader and expositor.

This chronology does not diminish Tao’s role. His familiarity with the problem makes his reaction unusually informative. He knows why earlier approaches stalled and which features of a new argument deserve attention.

His digestion also protects against a common failure in AI mathematics. A formal certificate can circulate faster than any specialist can understand its underlying idea. Tao slows that process down by reconstructing the proof in conventional language.

That reconstruction creates an independent intellectual test. If the proof can be reorganized into an elementary human argument, then its value extends beyond successful compilation.

It also reveals a possible future role for senior mathematicians. They may spend more time selecting machine outputs, identifying their conceptual core, relating them to literature, and converting them into reusable theory.

That work is not clerical. Choosing the right abstraction can require as much mathematical taste as discovering a proof route. It determines whether a result becomes knowledge or remains an isolated certificate.

The pressure falls most directly on workflows that treat natural-language plausibility as sufficient. Chat transcripts, model consensus, and confident explanations look weaker once verified formal artifacts become available.

Traditional publishing also faces pressure. A formal proof can be checked publicly before a journal completes review. Expert commentary can arrive within days, while conventional publication may take months.

Journals will still matter for priority, quality control, archival stability, and exposition. They may increasingly expect formal artifacts for results that arise through automated theorem proving.

AI laboratories face another pressure. Benchmark scores cannot fully demonstrate research usefulness. A credible open-problem result must reveal the problem source, search process, formal statement, verifier output, and expert audit.

Mathematicians face pressure too, but not simply through job replacement. They must learn how to formulate useful targets, inspect machine-generated definitions, and recognize when a verified proof contains a valuable idea.

The Sendov episode therefore challenges both AI boosterism and professional defensiveness. The machine’s contribution becomes credible because humans specified, inspected, and explained it. Human judgment becomes more effective because machines expanded and checked the search.

That interdependence is the central reversal. Better verification does not remove mathematicians from the process. It changes where their scarce attention produces the most value.

Three Signals Will Decide What This Result Means

Independent reproduction, a stable human proof, and reuse of the method will determine whether Sendov becomes a landmark or an isolated success.

The first signal is independent formal reproduction. Lean specialists should obtain the source, compile it in a documented environment, and inspect every nonstandard assumption.

A successful independent build would strengthen the claim that the certificate is portable rather than tied to one private setup. Discovery of a specification mismatch would weaken it immediately.

The review should publish the exact theorem statement and dependency list. That would let specialists compare the formal result directly with the classical formulation instead of relying on summaries.

The second signal is a stable, citable mathematical manuscript. Tao’s digestion provides an important bridge, but the field still needs a complete account with definitions, lemmas, references, and attribution.

If specialists can teach the argument and reproduce its key steps without the original AI session, the result becomes part of ordinary mathematics. If the proof remains understandable only through a large formal file, its scholarly influence will be narrower.

A conventional manuscript would also clarify contribution boundaries. It should state what Mazur supplied, what AI systems generated, what Lean checked, and what Tao’s later exposition added.

The third signal is methodological reuse. Researchers should test whether the same proof architecture can settle stronger variants, simplify earlier degree-specific results, or expose new inequalities for polynomial critical points.

The Phelps-Rodriguez strengthening is an obvious test because it sharpens the geometric relationship behind Sendov’s statement. Progress there would show that the method captures structure rather than one fortunate contradiction.

Reuse would also answer an important AI question. Did the workflow discover a transferable mathematical idea, or did it navigate a single formal search space successfully?

A transferable method would strengthen the case for AI as a research collaborator. An isolated certificate would still be valuable, but it would offer weaker evidence about general mathematical reasoning.

For developers, the lesson is that verifier-backed outputs deserve a different category from ordinary model responses. Systems should expose assumptions, dependencies, failed branches, and reproducible artifacts rather than only polished answers.

For researchers, the lesson is to preserve the full evidence chain. A conjecture’s wording, formal encoding, generated proof, compiled certificate, and human exposition should remain linked.

For knowledge workers, the broader pattern is equally relevant. AI output becomes more trustworthy when an external system can test it against explicit rules. Mathematics offers an unusually clean version of that principle.

The sendov conjecture now stands at the center of that transition. The best current description is an AI-assisted, Lean-verified proof announced by Lech Mazur and seriously analyzed by Terence Tao.

Calling it “AI solved mathematics” loses the most informative part of the event. The result matters because generation, verification, and human understanding were separated, then connected through auditable artifacts.

The next step is concrete: watch for independent Lean builds, a durable scholarly manuscript, and new theorems using the same mechanism. If all three arrive, this proof will mark more than the end of a 67-year problem.