AI Really Did Disprove an 80-Year-Old Conjecture. Physics Is a Different Problem.

For eighteen months the coverage of artificial intelligence and mathematics has run ahead of the record. This investigation tested the specific claims, in order, against primary documents. Of the eight headline claims in our own assignment, five held up with corrections, one held only in part, and two failed as stated.

The short version:

• Two famous conjectures really did fall, and the evidence is unusually strong. Erdős’s unit-distance conjecture, posed in 1946, was disproved in May and now carries a nine-author, human-verified companion paper; a leading mathematician said he would have accepted the digest for the Annals without hesitation. The Jacobian conjecture was refuted in dimension three and higher in July, with a full machine-checked proof in Isabelle.
• Sendov’s conjecture was resolved with an AI tool, then digested by Terence Tao and re-formalised at one-sixth the length.
• But the “AI solves Erdős problems” story is inflated. In the most careful audit available, 8 of 13 “solved” problems were rediscoveries of existing literature. The field’s own proudest early moment — the first fully autonomous LLM solution of an Erdős problem — collapsed within hours because Erdős had settled it in 1977.
• OpenAI’s Navier–Stokes claim is not a clean win. Clay has not awarded anything and calls the process “deliberately unhurried”. The result is disputed on credit. And an unreviewed October preprint argues that the machine-checked Lean proof does not correspond to the human-readable proof.
• Physics is not next — not in the sense the phrase implies. The verified AI achievements in physics are engineering: better turbulence closures, operational weather forecasts, error correction on real quantum hardware. No AI-designed material has been synthesised and independently confirmed as new to science. Quantum gravity, dark matter and black-hole information show no verified contribution at all.

The pattern across all of it is consistent, and it is the most useful thing in this article: AI has become extraordinarily good at producing results inside frameworks humans built. It has not yet demonstrated the skill that produces new frameworks — posing the right question, or inventing the principle that makes a hard thing easy. The mathematicians said this themselves, in writing, in September.

Part 1 — What AI has actually solved

The result that deserves the headline: Erdős’s unit-distance conjecture

The question is elementary to state. Put n points in the plane. Let g(n) be the maximum possible number of pairs that are exactly one unit apart. Erdős conjectured in 1946 that g(n) grows only slightly faster than n — that the humble square grid is essentially as good as it gets.

It is not. On 20 May 2026 an internal OpenAI general-purpose reasoning model produced an infinite family of point configurations that beats the grid by a polynomial factor. The same day, Will Sawin gave an explicit version of the bound. Within hours of the announcement, nine mathematicians — Noga Alon, Thomas Bloom, Tim Gowers, Daniel Litt, Will Sawin, Arul Shankar, Jacob Tsimerman, Victor Wang and Melanie Matchett Wood — had posted a digested, human-verified version of the argument.

That companion paper is why this result stands up where so many do not. Gowers’s assessment, as quoted in coverage of that paper, was that if a human had submitted it to the Annals of Mathematics he would have recommended acceptance without hesitation.

Why it matters beyond the puzzle: the construction used algebraic number theory — number fields of large degree and small discriminant, via a Golod–Shafarevich criterion — to answer a question that looks like elementary plane geometry. The method is the interesting part, and it is human mathematics, applied by a machine to a place nobody had thought to apply it.

Two corrections the coverage missed. First, do not confuse this with Erdős’s distinct distances problem, which is a different question and remains open. Second, the construction is asymptotic and inexplicit: mathematicians have noted the structure cannot sensibly be drawn on paper even for a small number of dots. This is a proof about growth rates, not a picture.

The Jacobian conjecture — refuted, but only where it was refuted

Statement: if a polynomial map from complex n-space to itself has a Jacobian determinant that is a nonzero constant, must it have a polynomial inverse? Keller asked in 1939. It is Smale’s sixteenth problem. It has attracted false counterexamples for decades, which is exactly why it is a useful test of the hype.

On 20 July 2026 Levent Alpöge announced an explicit map of degree 7 that settles it — in three dimensions and above. The certificate is two finite, exact, checkable identities: the Jacobian determinant equals −2 identically, and three distinct rational points map to the same output. That is the whole thing. You can check it by hand.

The verification is unusually complete: an Isabelle/HOL entry in the Archive of Formal Proofs, dated 20 July 2026, proves the partials, the determinant, the three-point collision and non-invertibility in every dimension at least three — and explicitly makes no claim about dimension two. Tao published a digestion on 21 July. Independent exact-arithmetic checks followed.

The correction: the Jacobian conjecture is not disproved. The plane case — n = 2, the famous form — remains open. Write “refuted in dimension 3 and higher”, or you have written something false.

The human contribution here is worth naming precisely: Akhil Mathew suggested the problem. Alpöge directed the search. Tao did the work of turning a miracle into an explanation. Anthropic’s Claude (Fable) is credited with work leading to the map. This was not a model handed a conjecture and returning a proof.

Sendov’s conjecture — resolved, then digested

Sendov’s conjecture dates to 1959. Given a complex polynomial of degree at least 2 whose zeros all lie in the closed unit disk, every zero lies within distance 1 of some zero of the derivative. It had been proved for degree below 9 and for sufficiently large degree, leaving an awkward middle.

Lech Mazur used an AI tool to resolve it for all n, with the proof verified in Lean. Then something more interesting happened. As Tao wrote on 12 August, the AI-generated proof was not in publication-ready human form, and digesting it took him several days with heavy AI assistance. The digested argument turned out to be remarkably elementary — the deepest input is Maclaurin’s inequality — and it also yields the full Phelps–Rodriguez strengthening. Tao then re-formalised the whole thing in Lean at roughly 15,000 lines, against about 90,000 for the original.

That ratio is the single most informative number in this article. Machine proofs are not yet explanations. A human had to find the short path, and the short path was six times shorter.

The Erdős problems — real, but this is where the story inflates

“AI is solving Erdős problems” is true and almost entirely misleading as usually told, because it merges at least four different things.

Google DeepMind’s AlphaProof Nexus is the rigorous case. An agent alternating language-model proof generation with Lean compiler verification autonomously resolved 9 of 353 formally stated open Erdős problems, and 44 of 492 OEIS conjectures, at a few hundred dollars per problem. It was published as a peer-reviewed research article in Science on 8 October 2026, with a commissioned Perspective. Named examples include a problem of Erdős and Sárközy from 1970; Science’s editor summary says two of the nine had been open for 56 years. This is real, reviewed, and modest in scope — it works because those 353 problems had already been formalised.

A 24-author Gemini case study is the deflationary counterweight, and it is the most honest document in the field. Screening 700 conjectures labelled open, it found 13 had been addressed — of which only 5 were genuinely novel autonomous solutions, and 8 were rediscoveries of existing literature. Its own conclusion: the “open” label reflected obscurity, not difficulty. It also flags literature-search failure and subconscious plagiarism as live risks.

OpenAI’s internal model, Astra, is credited by the community database with disproving two Erdős problems and settling a third. One of those, on multicolour Ramsey numbers, carries an independent seven-page human exposition by Rob Morris hosted on the database — genuine independent verification. The other two do not.

And the cautionary tale. The first widely celebrated case of a large language model fully autonomously resolving an Erdős problem was announced on Christmas Day 2025 and retracted within hours, when a reader pointed out that Erdős himself had settled it in 1977.

The database that tracks these problems lists 1,221 problems in total as of 10 November 2026. I could not retrieve a citable page that reproduces its per-status breakdown, so no breakdown is published here. [SOURCE — erdosproblems.com, retrieved by the research desk]

The Riemann hypothesis — no, and the company says so

This is the claim most likely to be overstated in aggregation, so precisely: the Riemann hypothesis was neither proved nor disproved.

What happened is bounded and real. On 10 August 2026 Anthropic published a research note describing an unreleased research version of Claude raising the proven lower bound on the proportion of non-trivial zeros of the zeta function that lie on the critical line — from 41.6% to 67.2%, unconditionally. The method combined three known human lines of work into a rank inequality. Two Anthropic mathematicians validated it; Brian Conrey and Dan Goldston examined it at short notice. Anthropic’s own caveat, in its own words, is that it does not expect these techniques to lead to proving the Riemann hypothesis.

Status: awaiting review. No journal version, no third-party formal certification. Every number comes from the company’s own note. It is a genuine improvement in a related bound. It is not progress on the hypothesis in the sense the headline implies.

Matrix multiplication — a fourth decimal place

Also true, also smaller than it sounds. In August 2026 a ten-author paper improved the upper bound on the matrix multiplication exponent to ω < 2.371177, from 2.371339. The gain is about 0.00016. The method was to reformulate a core optimisation so it could be solved in a larger setting, then use machine learning to design the optimiser and AlphaEvolve to refine it.

Separately, and more interestingly, AlphaEvolve found in 2025 a way to multiply two 4×4 complex matrices in 48 scalar multiplications — the first improvement on Strassen’s 49 since 1969.

Do not describe either as an AI breakthrough in matrix multiplication. The honest sentence: AI helped shave the fourth decimal place off the best proven exponent, inside a human-designed framework; and separately found a better small-matrix algorithm. Neither produces faster code at real matrix sizes.

One more, because it belongs in Part 5

On 1 August 2026 OpenAI published ten results attributed to an internal version of Astra, with Lean certificates for each, across sphere packing, codes, non-sofic groups, Connes’s rigidity conjecture, circuit complexity, quantum parallel repetition, the closest vector problem, Ehrhart’s volume conjecture, multicolour Ramsey numbers and extremal numbers.

The mathematics appears to stand. The framing does not. Scientific American reported on 6 August that Stephen Miller of Yeshiva said the sphere-packing result hinges on an argument published in a 2016 paper of his without credit, and that the pattern “points to research misconduct”. Francesco Fournier-Facio of Cambridge found the non-sofic-group construction recombines ideas from 2016 and 2019 papers — and then independently reconstructed it, which means the construction is sound while the novelty claim is not. OpenAI told the magazine it takes responsibility for correctness, and revised an earlier claim that the problems had seen no progress for at least a decade.

That is the whole verification problem in one episode: Lean proves the argument compiles. It does not prove the result is new.

Part 2 — Navier–Stokes: the million-dollar claim

What the equations actually say

The Navier–Stokes equations are Newton’s second law applied to a fluid treated as a continuous medium. Water in a pipe, air over a wing, blood in an artery. Written down in the nineteenth century; still not understood.

The question that carries a Clay prize is this: start a fluid at rest, smooth and quiet. Can it, entirely through its own internal motion, develop infinite speed in finite time? A “singularity” means the model predicts something a real fluid cannot do — infinite velocity. If that can happen, the equations are not a description of reality everywhere; they are a description that breaks.

Viscosity is the reason this is surprising. Viscosity smooths things out. It is a friction term, constantly draining energy from small eddies. The intuition is that a fluid cannot manufacture infinite speed out of nothing but its own swirling, because friction is always taking energy away.

The expected mechanism is subtle, and OpenAI’s own description is unusually good: a vortex, spinning and spiralling inward, getting thinner and longer “like spaghetti”. The central region shrinks while it speeds up, and does so in such a precise balance that the total energy stays finite — as the laws of physics require. Everything has to blow up together in exactly the right way, with the terms for acceleration, pressure, motion and viscosity becoming enormous while cancelling out.

Leray established in 1934 that generalised, or weak, solutions exist. Whether the smooth ones stay smooth has been open for roughly ninety years.

What OpenAI claimed, on 8 September 2026

OpenAI published a post titled On the Navier–Stokes Millennium Prize Problem, a paper, and a Lean formalisation. Its claim, in its own words: its system produced “an analytical proof and a Lean formalization that an initially smooth fluid at rest can develop a singularity in a finite time”, and that this “resolves the Navier–Stokes Millennium Prize problem by establishing statement ‘C’ (and also ‘D’) in the official Millennium Prize formulation”. [VERIFIED — retrieved, openai.com, 2026-09-08]

Three details from that post matter and were widely dropped:

1. The fluid has a smooth external force applied to it. OpenAI states this plainly: “The fluid has a smooth force applied to it, and its energy remains finite through the entire dynamics, from rest to the formation of the singularity.”
2. The proof was produced by about 10,000 concurrent agents, in a system set up after OpenAI “heard rumors that two Millennium Prize problems had been resolved” on 1 September.
3. It is a disproof of smoothness, not a proof of well-behavedness. Statements C and D are the breakdown alternatives.

On that first point I want to correct a natural assumption, including my own. Under Fefferman’s official formulation the breakdown alternatives explicitly allow a smooth external force; it is alternatives A and B — global smoothness — that are posed with no force at all [SOURCE — the A/B versus C/D distinction rests on a secondary explainer that quotes the official statement; Clay’s own PDF could not be text-extracted on this station]. So the forcing is not a loophole. What follows from that is narrower, and is the thing to report accurately: the result settles the breakdown direction. It does not decide the smoothness direction. A world in which singularities can be triggered by a well-behaved force is still a world in which unforced flows might always stay smooth.

One clarification a critic will reach for, because the same post contains a line that can look like a contradiction. Under the heading Concurrent work OpenAI writes: “In the Euler case, Alpöge and Buckmaster proved a result with external forcing, while OpenAI’s system proved a result without external forcing.” That sentence is about the Euler equation — the inviscid special case, Navier–Stokes with the viscosity term set to zero — and OpenAI links a separate Euler proof paper from the same post. It is a comparison with Alpöge and Buckmaster’s forced Euler result. The sentence quoted earlier, about a fluid with a smooth force applied, is about Navier–Stokes and statements C and D. Two different equations, two different papers: the two statements are not in conflict. [VERIFIED — both lines retrieved, openai.com, 2026-09-08; the “without external forcing” sentence sits in the “Concurrent work” section]

That reading is reinforced from the other direction. A preprint by Constantin, Ignatova and Vicol dated 17 September 2026 is reported to argue that OpenAI’s method does not extend to statements (A) and (B) — the unforced case. [SOURCE — reported in the Wikipedia article “Navier–Stokes existence and smoothness”, citing that preprint; the preprint itself was not retrieved for this article]

Clay’s answer, and what it does not say

On 11 September the Clay Mathematics Institute published a statement. It congratulated nobody and awarded nothing. It said the Institute “shares in the excitement of the global mathematical community as we contemplate the announcement that the Navier–Stokes problem has apparently been settled” — note apparently — and that “The process is deliberately unhurried, but we will provide updates.” [VERIFIED — retrieved, claymath.org, 2026-09-11]

Clay’s own page for the problem now carries a badge reading “Active”. Not solved. Not unsolved. [VERIFIED — retrieved, claymath.org/millennium/navier-stokes-equation]

Its president, Martin Bridson, has been reported as stressing that the Institute requires a solution in the form of peer-reviewed publications, which would then be scrutinised further before any prize is awarded. OpenAI, for its part, has said it does not plan to claim the award.

So: the Millennium Prize problem has not been officially solved. The prize has not been awarded. Any article saying otherwise is wrong. As of this writing the status is: claimed, published as a preprint and a Lean repository, criticised in detail, disputed on credit, and awaiting a process that describes itself as unhurried.

The credit dispute

About twelve hours before OpenAI’s announcement, the mathematician Tristan Buckmaster published a statement on behalf of himself and, in part, Levent Alpöge. Their recent work on the Euler equations — a special case of Navier–Stokes — had, he said, been passed to OpenAI days earlier and may have shaped the prompts used for the agents.

OpenAI denies direct use, and Sam Altman has said the approaches “appear to be different”. OpenAI’s researcher Sébastien Bubeck disputed parts of Buckmaster’s account publicly and apologised for his choice of words. Buckmaster separately raised whether unpublished drafts he had entered into OpenAI’s Codex coding tool could have influenced the model; OpenAI said an internal investigation found those inputs could not have done so, and that no user inputs after 3 July could have influenced the system.

Reporting on the episode describes a 6 September meeting at which Buckmaster was offered two options — report his own Euler result first, or rewrite OpenAI’s proof as sole or lead author — and says he declined both because Alpöge would be excluded from authorship of work they had initiated. It also reports that Alpöge is employed by Anthropic, OpenAI’s rival.

I should be explicit about the limits of my sourcing here. The primary documents in this dispute are a statement on Buckmaster’s departmental website and posts on social media, and I did not retrieve those directly. What I can verify is that the dispute is real, that it is reported in detail by the Financial Times, The New York Times, Fortune, Science and Nature, and that OpenAI’s post contains a section headed “Concurrent work”. Readers should treat the mechanics of the dispute as reported rather than established, and read the primary statements.

The finding that actually matters: the Lean proof does not match the paper

On 6 October 2026, Alexander Bastounis and Fabian Circelli of King’s College London and Cambridge, and Anders Hansen of Cambridge, posted a preprint with a title that gives the game away: Navier-Stokes lost in translation: Why Lean verification of AI autoformalisation does not guarantee correct natural language proofs.

Their finding, from the abstract: “we show that the formalised Lean proof does not correspond to the NL proof of blow-up of solutions to the Navier–Stokes equations.” [VERIFIED — retrieved, arXiv:2610.08144, 2026-10-06]

King’s College London’s own release is careful about what this does and does not mean, and I want to preserve that care. It says the finding “does not yet establish that OpenAI’s original proof is incorrect, or that its claimed mathematical result is wrong”, but that it “suggests that the written proof and its computer-verified version make different mathematical claims”. In one example, the written proof makes a claim stronger than the corresponding claim in the computer-checked version. [VERIFIED — retrieved, kcl.ac.uk, 2026-10-27]

The deeper result is not about OpenAI at all. To translate human mathematics into a formal language faithfully, you must resolve ambiguity, and the authors show that this problem sits arbitrarily high in the Solvability Complexity Index hierarchy — informally, harder than any computational problem including the halting problem. There is no general automatic procedure for doing it.

Two caveats, both important. The preprint has not been peer reviewed; King’s College London says so in the release. And it is a critique of the translation, not a refutation of the theorem.

But the consequence is what the whole feature turns on. An article that says “the proof is Lean-verified, therefore it is correct” is making an error. A machine-checked proof establishes that a formal object follows from its axioms. It does not establish that the formal object is the paper you published.

Bastounis put it plainly: “OpenAI’s struggle to accurately translate its proofs into a machine-checkable language highlights a broader challenge: verifying the correctness of the output of an AI remains exceptionally difficult.”

What this case is a case of

Here is what I think the Navier–Stokes episode actually demonstrates, and it is not that AI cannot do mathematics.

It is that the bottleneck has moved. Producing a candidate proof is no longer the scarce thing. Translating it faithfully, checking it against the literature, assigning credit, and digesting it into a form a human can learn from — those are now the scarce things. Every one of them requires people.

Terence Tao wrote, one day before the announcement, in a post about the Euler work that preceded it: “the actual solving of these problems is only a proxy goal for the primary goal of developing mathematical understanding and insight. Without such understanding, even a problem as infamous as the Navier-Stokes regularity problem [is] of far less intrinsic significance to mathematics than is sometimes promoted in popular media.” [VERIFIED — retrieved, terrytao.wordpress.com, 2026-09-07]

That sentence is the thesis of this article. It was written before the announcement, by the person best placed to judge, and it has not been answered.

Part 3 — The great unsolved problems, and what AI has done to them

A note on method: Clay’s own site badges each Millennium problem, and I verified that Navier–Stokes now reads “Active” rather than unsolved. The other problems below remain open; where AI progress is claimed, I have graded it.

Problem	In plain English	Why it matters	Status (Nov 2026)	Verified AI progress
Riemann Hypothesis	The zeros of a certain complex function all sit on one vertical line. If true, it tells us how the primes are distributed.	The primes are the atoms of arithmetic; their distribution underpins cryptography and much of number theory.	Open since 1859	Indirect only. A proven lower bound on the proportion of zeros on the line rose from 41.6% to 67.2% (Anthropic, company-reported, awaiting review). The hypothesis itself is untouched.
P vs NP	If a problem’s answer can be checked quickly, can it always be found quickly?	It is the difference between a lock that can be picked and one that cannot. Modern cryptography assumes they differ.	Open since 1971	No meaningful verified progress. No AI result addresses the known barrier theorems.
Yang–Mills and the mass gap	Prove that a class of quantum field theories exists mathematically, and that its lightest particle has positive mass.	It would put the Standard Model’s foundations on rigorous footing.	Open since 2000	Claimed, unread, disputed. OpenAI announced in October that its models produced an existence proof, formalised in Lean, but withheld the physics materials pending an external working group. Critics note the mass gap is not addressed and the word “existence” is undefined in the release.
Hodge Conjecture	Certain geometric shapes that arise from equations should be built from simpler pieces called algebraic cycles.	It links topology to algebra — two ways of describing shape — with consequences across geometry.	Open since 1950	No meaningful verified progress.
Birch and Swinnerton-Dyer	The number of rational solutions to an elliptic curve equation is governed by a related analytic function.	Central to modern number theory and used in cryptography.	Open since the 1960s	No meaningful verified progress.
Navier–Stokes (cross-reference)	Can a smooth fluid develop infinite speed on its own?	It decides whether the equations describing water, air and blood are universally valid.	Clay badge: “Active”. Not solved; prize not awarded.	Disputed and unresolved. OpenAI’s September claim covers the breakdown direction with a smooth force. Credit is disputed; the Lean proof does not match the paper per an October preprint.
Goldbach’s Conjecture	Every even number above 2 is the sum of two primes.	A basic statement about the building blocks of arithmetic that we cannot prove.	Open since 1742	No meaningful verified progress. Verified numerical ranges are computation, not proof.
Twin Prime Conjecture	There are infinitely many prime pairs differing by 2.	Same family: the deepest simple questions about primes remain open.	Open; best results bound gaps, not twins	No meaningful verified progress.
Collatz Conjecture	Repeatedly halve even numbers and triple-plus-one odd numbers; do you always reach 1?	Elementary to state, apparently intractable — a benchmark for what proof can and cannot reach.	Open since 1937	No meaningful verified progress. Verified ranges are computation.
abc Conjecture	A relation between a number’s prime factors and sums of nearby numbers.	Would imply many other results at once.	Disputed status after Mochizuki’s claimed proof	No meaningful verified progress.
Smooth 4-dimensional Poincaré	The one dimension where the Poincaré conjecture remains unsettled in the smooth case.	Dimension four is where topology behaves most strangely.	Open	No meaningful verified progress.

The honest pattern: AI has made verified progress on problems that were already formalised, or that decompose into computation. Where the difficulty is conceptual rather than technical — where the barrier is a missing idea rather than a missing calculation — there is no verified AI result at all.

Part 4 — Is physics next? A four-rung ladder

This section is where coverage most reliably goes wrong, so it is organised by an explicit standard. Every claim gets exactly one rung:

• (A) Mathematical consistency — a proof or formal argument. No experiment, and possibly no experiment is possible.
• (B) Simulation — reproduced in code. No experiment.
• (C) Experimental prediction — a specific testable claim, tested against real measurements.
• (D) Independently confirmed discovery — validated experimentally by an independent group.

Most “AI discovers physics” stories collapse A through C into D. They should not.

(D) What has actually been confirmed — a short list

Quantum error correction on real hardware. In work published in Nature on 8 July 2026, Google Quantum AI applied reinforcement learning to live control of a Willow superconducting processor, reporting roughly 20% additional logical-error suppression on top of exhaustive human calibration, and record performance including a distance-7 surface code. This is the strongest AI-in-physics result available, because it left the laptop and touched hardware.

The caveat is real: the same collaboration designed, ran and evaluated the experiment, no independent replication was found, and “record” is the paper’s own claim.

One compound, one property, one collaborating lab. MatterGen generated a structure that was synthesised and whose measured property came within 20% of the target — Nature 639, 624–632 (2025), Microsoft with a synthesis partner in Shenzhen. The authors call it proof of concept. One material is not a materials revolution, and the synthesis was not independent.

A plasma, in one laboratory. A physics-tailored neural network inferred non-reciprocal interparticle forces from 3D particle tracking in a dusty plasma with better than 99% accuracy, and showed some standard theoretical assumptions were wrong (PNAS, 2025). The lead author’s framing was that they showed AI can discover new physics. What the paper shows is that AI can infer a force law inside a known framework, in one system, in one lab, without independent replication.

(C) What works in the real world, and is engineering

Turbulence closures. A physics-informed neural network closure trained on towing-tank and wind-tunnel particle-image velocimetry for cylinder flow significantly improves predictions of mean flow and turbulence statistics over conventional closures, embedded in a real solver (Journal of Fluid Mechanics, volume 1034, April 2026). This is the genuinely impressive work, and it is a better engineering correction to an approximate model — not a new theory of turbulence. Nobody has produced one.

Weather forecasting. ECMWF’s machine-learned AIFS has been operational since February 2025; its 2026 version improves overall skill by 4–6%, and the v2 release added data-driven wave and snow-cover forecasts that outperform the physics-based model on wave skill. Operational adoption against observations is the strongest validation a forecast model can get. It is also not a discovery.

Cosmology. A hybrid physics-plus-machine-learning analysis of Dark Energy Survey data extracted about three times more cosmological information than traditional statistics, giving the most precise weak-lensing-only measurements to date and easing a long-standing tension with the cosmic microwave background. The machine learning is an estimator. It did not find dark matter.

(A) Theoretical results — real, checkable, and not physics yet

A gluon amplitude formula. GPT-5.2 Pro conjectured a closed-form expression for single-minus gluon tree amplitudes, after humans had computed small cases by hand; a scaffolded internal model produced a formal proof, and the formula satisfies Weinberg’s soft theorem and solves the Berends–Giele recursion (arXiv:2602.12176). Nima Arkani-Hamed’s assessment, on OpenAI’s own page: finding a simple formula “has always been fiddly, and also something that I have long felt might be automatable by computers”.

Note the scope: half-collinear configurations, in Klein space or complexified momenta. Narrower than “AI does impossible gluon physics”.

A nine-loop amplitude. Claude computed the six-particle scattering amplitude in planar N=4 super Yang–Mills at nine loops, one beyond the previous record, answering a public challenge, checked with amplitudes researcher Lance Dixon. Published as a guest post on Anthropic’s site, with a disclosure that Anthropic paid the author and gave feedback on drafts. The method was established bootstrap techniques — known technique plus unusual persistence, not a new method.

The withheld Yang–Mills claim. On 26 October OpenAI said its models produced a mathematically rigorous existence proof for certain four-dimensional quantum field theories including QCD, formalised in Lean, but withheld the physics materials pending an external working group chaired by David Gross. An advisory group of mathematicians formed to advise OpenAI on how to release results publicly called for immediate release. Peter Woit’s objection is precise and unmet: the release does not say what “existence” means, and the mass gap is not addressed. [SOURCE — openai.com and Woit, both listed in Sources]

Because the document has not been read by anyone outside the process, it does not belong on the ladder. It is a claim.

(B) Simulation — and the three areas with nothing

Quantum gravity: no verified contribution. What exists is inverse-problem engineering inside holography — feed a network boundary data from a known gravitational solution and it reconstructs the bulk (arXiv:2609.18566). The answer is known in advance; the exercise is numerical method.

The most recent survey of AI for physics discovery is blunt about the ceiling. Shalyt, Regev, Soljačić and Kaminer, in Can AI Follow In Einstein’s Footsteps?: “If this trend continues, AI would become extraordinarily good at prediction but may struggle to ever propose its first serious contender to quantum gravity or other paradigm-level theories.” [SOURCE — arXiv:2607.27794] The review names a missing skill rather than a missing machine: it highlights “the ability to pose the right questions or invent the right principles” as the critical gap.

Dark matter and dark energy: no verified contribution to what either is. No candidate particle identified, no modified-gravity proposal tested. Everything retrieved is statistical methodology.

Black hole information: no verified contribution. The one relevant paper treats the Page curve as an analogy to interpolation thresholds in quantum machine learning — an observation about mathematical similarity, explicitly not a resolution.

Finally, a warning about the last 48 hours. Several aggregators are reporting that a Claude model solved two open black-hole problems. I could find no primary document — no Anthropic publication, no preprint. Under our standard that claim stays off the ladder entirely until someone produces the paper.

The materials correction, because it is the clearest documented failure

If you want one story that shows why rung (D) matters, take A-Lab.

The autonomous laboratory’s headline was that it created dozens of new materials in days. In January 2026 Nature published an Author Correction: the novelty claims, the authors wrote, “were subject to misinterpretation — their intention was to indicate that the materials were new to the prediction platform, not necessarily new to science.” Manual re-analysis confirmed the platform’s conclusion in 36 of 40 reported successes; four were inconclusive; one compound was found to have been in the training data.

New to the platform. Not new to science. That distinction, stated by the authors themselves, is the difference between a discovery and a database update, and it is the distinction most coverage of AI in science erases.

DeepMind’s GNoME figure — that outside labs “independently created 736 of these new structures experimentally in concurrent work” — is the company’s own characterisation of concurrent work, and partly rests on that same A-Lab set. I could not trace any independent synthesis paper attributing a specific new compound to a specific GNoME prediction.

No AI-designed material has, in anything I could retrieve, been synthesised by an unaffiliated laboratory and independently confirmed as new to science with replicated characterisation.

What physicists say

Anastasia Volovich of Brown, on the gluon result: it “goes a long way towards establishing trust that frontier AI systems can reliably perform difficult calculations”. And then, in the same interview, the sentence that should be quoted more often than it is: “It remains to be seen whether AI systems will be able to have genuinely novel insights, invent genuinely new methods or find genuinely new structures.” [SOURCE — ibm.com/think, listed in Sources]

Woit, on the withheld Yang–Mills announcement: “From this press release, it’s hard to know exactly what they’re claiming to have proved.”

So: is physics next?

Not in the sense the question usually carries. Physics is not waiting for a faster calculator. The unsolved problems of physics are unsolved because we lack a principle, not because we lack computational throughput — and the one capability AI has not demonstrated is producing principles.

What physics is getting, and should be glad of: better models of turbulence, better forecasts, better error correction, better estimators for telescope data. These are real, deployed, and measurably better than what came before. They are rung (C). Calling them discoveries is how the field loses the public’s trust when the next real one arrives.

Part 5 — If AI becomes a scientific discoverer

Assume the trajectory continues. What actually changes?

The verification bottleneck becomes the rate-limiting step, and it is a human bottleneck. The Navier–Stokes case is the template. A result appeared; nobody could confirm that the machine-checked version said what the paper said; the critique needed three mathematicians and a preprint; and the institution that awards the prize describes its process as “deliberately unhurried”. Compute scaled to ten thousand agents. Verification scaled to three people.

Credit assignment will get worse before it gets better. Two disputes are live in this article — the Navier–Stokes priority fight and the Scientific American reporting on OpenAI’s ten advances, in which a named mathematician alleges research misconduct over an unattributed 2016 argument. In both, the mathematics may stand while the attribution fails. Mathematical communities run on attribution: it is how careers, hiring and funding are allocated. A system that produces results faster than it can attribute them corrodes that mechanism whether or not any individual actor behaves badly.

The cost asymmetry is new and it is a problem. New Scientist reported OpenAI’s effort at around $15 million of compute; a researcher at the company confirmed publicly that it cost millions. Against that, a research mathematician’s grant is very small. Several mathematicians have objected that this converts mathematics from a contest of intelligence and diligence into a contest of capital. It also means the results will be produced by a handful of firms, on their own schedule, announced via blog posts.

That last point has an institutional edge. On 11 September 2026, twenty-five Fields Medallists — since grown — published a declaration, A Severe Misalignment of AI in Mathematics. Its central claim is not that AI is bad at mathematics. It is that the profession’s goals and the companies’ goals are different: “solving problems is only a tool and proxy for achieving the primary goal of conceptual understanding and insight”, and “the mass production at faster and faster pace of ‘true/false’ statements could destroy fertile ground instead of breathing life into new ideas”. [VERIFIED — retrieved, mathandai.org, DOI 10.5281/zenodo.22737750]

They also name the mechanism. Results are “announced in a rush, leaving no time for a proper writeup, the isolation of new methods and ideas, and citing relevant previous work of others. As in all creative professions, this raises severe attribution and plagiarism questions.”

That was signed by essentially the entire leadership of the field. It is not a fringe complaint, and it aligns exactly with what the A-Lab correction and the Scientific American reporting show in practice.

What does not change: understanding is still the product. Tao’s line again — solving is a proxy for understanding. The Sendov case is the optimistic version: an AI found a 90,000-line proof; a human found the 15,000-line one; the short proof is the contribution to mathematics. If AI keeps producing long correct proofs and humans keep producing short explanatory ones, the division of labour is stable and productive. If the field stops funding the humans who do the shortening, it will accumulate verified results nobody understands — which is a strange kind of progress.

And the thing to be careful about. Every overstatement this investigation examined traced back to the same move: collapsing a bounded, conditional or company-reported result into an unbounded claim. A bound on the proportion of zeros became “Riemann solved”. A forced blow-up became “Navier–Stokes solved”. Materials new to a prediction platform became “new materials”. The mathematics of 2026 is genuinely remarkable. It does not need the inflation, and the inflation is what will make the public disbelieve the next real breakthrough.

What survived scrutiny

Testing the eight claims in our assignment:

Claim as given	Verdict
Erdős unit-distance conjecture disproved	Holds, and is the strongest result of the year. Companion paper with nine authors; explicit bound from Sawin. Do not merge with the distinct-distances problem, which is still open.
Jacobian conjecture counterexample	Holds for dimension 3 and above. The plane case remains open. Machine-checked in Isabelle.
Sendov’s conjecture progress	Understated in the brief — it was resolved. Then digested by Tao from about 90,000 to about 15,000 Lean lines.
2026 Erdős problem solutions	Partly holds, heavily inflated. DeepMind’s 9 are peer-reviewed in Science; a careful audit found 8 of 13 “solved” elsewhere were already in the literature.
Riemann zeta improvements	Holds, narrowly. 41.6% to 67.2% of zeros proven on the critical line. Company-reported, awaiting review. Not the hypothesis.
Matrix multiplication advances	Holds, trivially. ω improved in the fourth decimal place. AlphaEvolve’s 48-multiplication 4×4 (2025) is the more interesting result.
OpenAI Navier–Stokes Millennium Prize solved	Fails as stated. Claimed for the breakdown direction with a smooth force; Clay has awarded nothing and badges the problem “Active”; credit disputed; Lean proof shown not to match the paper (preprint, unreviewed).
Physics is next	Fails as stated. No verified AI contribution to quantum gravity, dark matter, dark energy or black-hole information. Materials claims were formally corrected. The real wins are engineering.

Methodology and limits

• Every factual claim above rests on a document retrieved during this investigation — preprints, journal pages, institutional statements, company research notes — not on search snippets. Full sources are listed with the Evidence File.
• Quotations from Nima Arkani-Hamed, Anastasia Volovich, Peter Woit, Justin Burton, Terence Tao, the Fields Medallists’ declaration, and OpenAI’s and Clay’s statements were taken from the pages cited.
• Where I did not retrieve a primary, I have said so. That applies to the mechanics of the Navier–Stokes priority dispute, whose primary statements are on social media and a departmental website; to Timothy Gowers’s assessment, quoted in coverage of the companion paper rather than from a document of his that I read; and to a KITP working-group statement on the Yang–Mills claim whose PDF I could not extract.
• I could not text-extract Fefferman’s official problem description from Clay’s PDF. The A/B versus C/D distinction as described here rests on a secondary explainer that quotes the official statement, plus OpenAI’s own claim. It should be checked against the PDF before publication.
• The October preprint criticising the Lean formalisation has not been peer reviewed; neither has the Anthropic zeta note, nor the matrix multiplication preprint.
• No human was interviewed for this article.

FOR HUMAN CONSUMPTION

Sources

1. OpenAI, “On the Navier–Stokes Millennium Prize Problem”, 8 September 2026. openai.com
2. Clay Mathematics Institute, “Navier-Stokes Announcement”, 11 September 2026. claymath.org
3. Clay Mathematics Institute, Navier-Stokes Equation problem page (status badge “Active”). claymath.org
4. Bastounis, Circelli, Hansen, “Navier-Stokes lost in translation”, arXiv:2610.08144, 6 October 2026. arXiv
5. King’s College London, “OpenAI’s Navier-Stokes proof called into question”, 27 October 2026. kcl.ac.uk
6. Terence Tao, “Finite time blowup with smooth forcing term for the incompressible porous medium, Boussinesq, and incompressible Euler equations”, 7 September 2026. terrytao.wordpress.com
7. Avila, Bhargava, Birkar, Deligne, et al., “A Severe Misalignment of AI in Mathematics”, 11 September 2026, DOI 10.5281/zenodo.22737750. mathandai.org
8. Terence Tao, “A digestion of the proof of Sendov’s conjecture”, 12 August 2026. terrytao.wordpress.com
9. Terence Tao, “A digestion of the Jacobian conjecture counterexample”, 21 July 2026. terrytao.wordpress.com
10. Isabelle Archive of Formal Proofs, “The Jacobian Conjecture Counterexample”, 20 July 2026. isa-afp.org
11. Alon, Bloom, Gowers, Litt, Sawin, Shankar, Tsimerman, Wang, Matchett Wood, companion paper on the unit-distance counterexample, arXiv:2605.20695. arXiv
12. Sawin, explicit bound for the unit-distance problem, arXiv:2605.20579. arXiv
13. Google DeepMind AlphaProof Nexus, Science 394(6820), 8 October 2026. science.org
14. Feng et al., Gemini case study of open conjectures, arXiv:2601.22401. arXiv
15. Anthropic, research note on the Riemann zeta function bound, 10 August 2026. anthropic.com
16. Dupont et al., matrix multiplication exponent, arXiv:2608.16884, 17 August 2026. arXiv
17. OpenAI, “Ten advances in mathematics”, 1 August 2026. openai.com
18. Scientific American, “OpenAI’s Latest Math Breakthroughs Commit Research Misconduct, Experts Say”, 6 August 2026. scientificamerican.com
19. Nature Author Correction, A-Lab, 19 January 2026. nature.com
20. Sivak et al., reinforcement learning for quantum error correction on Willow, Nature, 8 July 2026. nature.com
21. Guevara, Lupsasca, Skinner, Strominger, Weil, single-minus gluon amplitudes, arXiv:2602.12176. arXiv
22. OpenAI, “New result in theoretical physics”. openai.com
23. OpenAI, “Working with academic communities to build scientific understanding”, 26 October 2026. openai.com
24. Peter Woit, Not Even Wrong, on the Yang–Mills release. columbia.edu
25. Shalyt, Regev, Soljačić, Kaminer, “Can AI Follow In Einstein’s Footsteps?”, arXiv:2607.27794. arXiv
26. Journal of Fluid Mechanics 1034, A16, physics-informed RANS closure, 27 April 2026. cambridge.org
27. ECMWF, IFS Cycle 50r1 / AIFS v2 live announcement, 12 May 2026. ecmwf.int
28. Imperial College London, machine-learning weak lensing analysis of Dark Energy Survey Year 3 data, 2026. imperial.ac.uk
29. Microsoft/SIAT, MatterGen, Nature 639, 624–632 (2025). nature.com
30. Emory University, dusty plasma AI force-law inference (PNAS 2025). emory.edu
31. IBM Think, “AI is getting uncomfortably good at physics”, quoting Anastasia Volovich. ibm.com
32. Anthropic, “Yes, Claude can do nine loops”, 25 September 2026. anthropic.com