An AI-found counterexample to an 87-year-old conjecture can be verified by anyone in seconds. The commercial story labs build around such results cannot, and letting the checkable claim vouch for the unauditable one is verification arbitrage.
Here is a claim that does not ask you to trust the mathematician who posted it, the AI lab it credits, or a word of the coverage around it. Take the polynomial map that Levent Alpöge published on July 19, three polynomials in three variables:
Differentiate the three coordinates with respect to x, y, and z, assemble the three-by-three Jacobian matrix, and take its determinant. It simplifies to the constant -2, which is never zero. Then evaluate the map at three distinct points, (0, 0, -1/4), (1, -3/2, 13/2), and (-1, 3/2, 13/2). All three land on the same output, (-1/4, 0, 0). A map whose derivative is invertible at every point, and which nonetheless sends three inputs to one image, so it cannot have a polynomial inverse.
If you would rather not do it by hand, this settles it:
import sympy as sp
x, y, z = sp.symbols("x y z")
P = (1 + x*y)**3*z + y**2*(1 + x*y)*(4 + 3*x*y)
Q = y + 3*x*(1 + x*y)**2*z + 3*x*y**2*(4 + 3*x*y)
R = 2*x - 3*x**2*y - x**3*z
F = (P, Q, R)
assert sp.factor(sp.Matrix(F).jacobian((x, y, z)).det()) == -2
pts = [(0,0,sp.Rational(-1,4)), (1,sp.Rational(-3,2),sp.Rational(13,2)), (-1,sp.Rational(3,2),sp.Rational(13,2))]
imgs = {tuple(sp.simplify(f.subs({x:a,y:b,z:c})) for f in F) for a,b,c in pts}
assert imgs == {(sp.Rational(-1,4), 0, 0)}
Run it. The determinant is -2 or it is not. The three points collide or they do not.
They do.
That is a counterexample to the Jacobian conjecture in dimension three, a problem open since Ott-Heinrich Keller posed it in 1939. Alpöge posted the map and credited Anthropic's Fable model with finding it. Independent symbolic checks appeared within hours. I argued a few days ago that enterprise AI lives in a self-report economy , a market where every load-bearing number about whether the technology pays off was produced by a party with a stake in the answer, so every figure needs a discount before you can use it. This is the inverse of that. The artifact is verifiable on your machine, not theirs. The commercial story labs build around artifacts like it usually is not, and telling the two apart is the whole subject of this piece. When you cannot prove the product pays, you prove a theorem instead.
What actually happened, and what actually checks out #
The Jacobian conjecture is one of those statements that sounds like it has to be true. Take a polynomial map from complex n-dimensional space to itself. If its Jacobian determinant, the quantity measuring how the map stretches space locally, is a nonzero constant everywhere, then the map should have a polynomial inverse. Local invertibility forcing global invertibility. In one variable it is a triviality: a polynomial with a nowhere-vanishing derivative is linear, by the fundamental theorem of algebra, and a linear map is invertible. The conjecture claimed the same rigidity survives into every dimension. For eighty-seven years nobody proved it and nobody broke it.
Alpöge broke it in dimension three. His map has a constant Jacobian determinant, which makes it what algebraists call a Keller map, and three distinct points with real coordinates all map to the same place, which makes it not injective. A Keller map that is not injective is exactly the object the conjecture said could not exist. The verification is the entire point, so hold on it. The checks that went up within hours, a SymPy script, a Sage script linked from an answer on MathOverflow, confirm the determinant and the collision as identities of polynomials, not as samples taken at a few coordinates and extrapolated. An identity holds everywhere or fails, and this one holds. The collision uses only real numbers, so it certifies non-injectivity over the reals and the complex numbers at once. Nobody had to convene a committee. Anyone who doubted it could rerun the check before lunch.
One correction to the way the result is circulating. This is not a degree-four map. Its three coordinate polynomials have total degrees seven, six, and four, so it is a degree-seven map, and the tidy "degree four" that appeared in some early write-ups, mine included, is wrong. A second correction matters more, because it is the kind of overstatement the moment invites. The counterexample does not invalidate the classical theorems that reduce the general conjecture to cubic or homogeneous forms after adding variables. Those reductions are still correct. What changes is their job. Instead of reducing a conjecture someone hoped to prove, they now transform a known counterexample into other special shapes. Append unchanged coordinates to Alpöge's map and the determinant stays constant and the collision survives, so the conjecture is false in every dimension at or above three. Dimension two, oddly, remains open. The plane is still holding.
The map is elementary to check. The process that produced it is not yet equally transparent, and that seam runs through everything below. The public artifact establishes the mathematics. It does not, on its own, reveal the full prompt history, the search strategy, the count of failed attempts, the human interventions, or how much of the decisive move came from the model rather than its operator. The gap between an auditable artifact and a far less auditable account of how it was made is the central fact of the story.
A claim that audits itself #
So name the property precisely. Its evidence is public, portable, cheap to evaluate, deterministic, and sufficient to settle the exact claim being made. I call an assertion with those properties a zero-discount claim, and the word zero should be read narrowly. It does not mean every statement near the result deserves full confidence. It means the interested party's incentives give you no reason to discount the truth of the specific, executable claim. Anthropic could benefit commercially from the counterexample. Anthropic's incentives cannot change the determinant of the displayed map. The artifact has escaped the claimant, which is rare and which is worth a great deal.
Most claims reach us through institutions. Auditors, journals, regulators, rating agencies, inspectors, trusted intermediaries, each vouching for something the audience cannot check directly. A zero-discount claim travels with enough evidence that the claimant's reputation stops mattering. You do not have to ask whom to believe. You ask the machine a narrower question and get a decisive answer.
Mathematics is not the only field that can produce such artifacts. Software can ship with reproducible tests. Cryptographic results can carry machine-checkable proofs. A data analysis can include the full dataset and an executable notebook. But mathematics is one of the cleanest cases, because correctness can sometimes be compressed into an object whose verification is radically cheaper than its discovery. Finding Alpöge's map may have taken expert judgment, heavy model inference, a search across many candidate constructions, or all three. Checking it takes seconds. That asymmetry, expensive discovery and cheap verification, is what makes the result both impressive and trustworthy in the same breath.
The credibility channel #
Mathematics has become an unusually valuable credibility channel for the frontier labs, and the commercial reason is not hard to see. An enterprise buyer wants evidence that a model will improve a specific workflow under real organizational constraints, which means answering questions about reliability, integration cost, exception handling, employee behavior, security, governance, and the counterfactual performance of the process already in place. Those questions are expensive to answer and the answers are usually specific to one company. Vendors see their most favorable deployments and have little reason to publish the failures. Customers cannot disclose their internal data. Independent evaluators rarely get the whole system. Every commercial value claim is born into the self-report economy, pre-discounted.
A mathematical artifact has none of those disadvantages. It can be hard enough to command attention, compact enough to circulate online, and verifiable enough to survive hostile scrutiny. It delivers the one thing almost no commercial AI claim can: a result that impresses and that a skeptic can independently confirm. There is a capability story underneath this, and it deserves a sentence. As raw model capability converges across the frontier, a pattern this site calls Model Convergence Pressure , the differentiation between labs stops coming from the base model and moves to what can be staged around it, and a headline-grade theorem is differentiation theater of the highest order. It says: our model, specifically, did what yours could not.
And mathematics is not merely a zero-discount domain. It is the rare one where the checkable claim can also be hard. Most self-verifying claims are trivial, which is exactly why they are useless as credibility. I can prove to you that two plus two is four and you learn nothing about me. A famous conjecture falling is unfakeable and difficult at once, and that pairing is nearly unique to mathematics and formal proof. It is why chess and Go registered as milestones and a benchmark score never quite does. The win is legible to a skeptic and still means something.
None of which makes the achievement cosmetic. A model contributing materially to a counterexample against a famous conjecture is a real capability signal. It is a signal about one capability under one workflow. It is not an audited estimate of enterprise value, and the temptation is to set the two side by side anyway. The model helped resolve a famous problem, the pitch runs, therefore the model will transform your business. The first clause is backed by a public artifact. The second is a broad prediction about a different environment resting on a different evidence base. Putting them next to each other invites the credibility of the first to flow into the second.
That maneuver has a name: verification arbitrage. An organization earns trust in a domain where claims are cheap to check, then spends that trust in domains where its claims stay hard to audit. The theorem is the deposit. The sales narrative is the withdrawal. The move works because audiences treat credibility as a property of an institution, this lab produced something undeniable, so its other claims deserve more benefit of the doubt. Credibility should attach to the evidence behind each claim, not to the logo on it. A verified counterexample establishes that one artifact is correct. It does not establish that the same system will process an insurance claim, manage a regulatory exception, forecast demand, or replace an analyst. Those are separate propositions, and they need separate audits.
Four claims hide inside one headline #
A line like "AI disproves an 87-year-old conjecture" compresses several claims that carry very different evidentiary weight. Pull them apart, because the announcement is built to keep them fused.
The artifact claim is that the displayed map is a genuine counterexample. This is the strongest of the four. It is independently reproducible and it holds. Give it close to full weight.
The contribution claim is that Fable made a material contribution to finding it. Alpöge credited the model publicly, so there is good reason to believe it played a real role, and I do. But the full interaction record did not ship with the mathematics. The exact split of credit between researcher, model, prompting strategy, and surrounding tools is not reproducible from the public evidence. Give it serious but qualified weight.
The capability claim is that Fable can reliably discover important new mathematics. One result raises the probability that the model is useful for such work. It does not establish a rate, because the denominator is missing. We do not know how many problems were attempted, how many candidate constructions failed, how much expert screening was needed, or how often a comparable workflow produces nothing at all. Treat it as promising evidence awaiting a larger sample.
The generalization claim is that the model's mathematical success shows it can autonomously perform difficult knowledge work across industries. This is the weakest of the four by a wide margin. It leaps from a domain with exact verification, forgiving iteration, and expert supervision into domains with ambiguous objectives, shifting data, tangled organizational dependencies, and expensive errors. It is a commercial hypothesis wearing a proof's clothes.
Four claims, one story, four different amounts of backing. The reader who keeps them fused lets the certainty at the top of the ladder pay for the uncertainty at the bottom.
The asymmetry runs opposite to the money #
There is a second inversion under the first, and it is the one I find genuinely strange. In the self-report economy, the numbers that matter most are the hardest to check. A Forrester total-economic-impact study sits behind a paywall and rests on a composite company you are not allowed to inspect. A lab's internal eval is a figure with no reproducible method attached. The more commercially load-bearing the number, the more locked down it tends to be, which is the whole reason the discount has to exist.
The zero-discount claim runs the other way. It is free, public, and self-verifying, available to a graduate student with a laptop and no institutional access of any kind. Alpöge did not paywall the map. He could not have made it more checkable if he had tried.
Hold those two facts together and the shape of the problem sharpens. The field has learned to produce claims of perfect, free, public verifiability in the one domain where the stakes are lowest, which is pure mathematics, while the claims with real money riding on them, the ROI figures and the deployment results, stay private, expensive, and unauditable. The capability to verify exists. It has been aimed at the cheap question. Nobody has aimed it at the expensive one, because the parties who could are the parties who profit from the fog, which is where the last piece ended and which the Jacobian result, of all things, quietly confirms.
When the artifact cannot be gamed, the framing becomes the game #
There is a structural reason this keeps happening, and it extends a pattern this site already named. Benchmark Contamination is the observation that any evaluation an interested party can both produce and be scored on gets gamed in the end, which is why SWE-bench and GAIA saturated and lost their meaning, and why serious evaluation had to move to held-out data and live trajectories the model builders never see. The fix was to relocate measurement to ground the measured party cannot author.
Mathematics looks, at first, like the terminal version of that ground. You cannot contaminate a counterexample that did not previously exist, because there was no training set of Jacobian counterexamples to overfit. That is real, and it is why the zero-discount claim is genuinely special.
But the contamination does not die when it reaches that ground. It climbs one level, off the artifact and onto the frame around it, and the year's record shows exactly how.
Consider the 2025 International Mathematical Olympiad. OpenAI and Google DeepMind each reported gold-medal-level performance after their systems solved five of six problems. DeepMind went through the IMO's coordinated evaluation. OpenAI published earlier, on the strength of assessments from external medalists, saying it had permission from an IMO board member after the closing ceremony, and DeepMind criticized the timing because the organizers had asked labs to wait for a coordinated release, as Reuters reported at the time. The mathematical outputs could still be inspected. The disputed variables were priority, timing, process, and endorsement, none of which a proof-check touches. Epoch AI then published a post-mortem arguing the medal revealed less than the headline implied: the five solved problems sat within the range earlier systems could already handle, while the sixth was far harder than anything the systems had solved, so the result showed more reliability rather than a clean new frontier of creativity. The score was real. The reading of the score was contestable.
The Erdős problems make the same point more sharply. In October 2025 a claim spread that GPT-5 had solved ten open Erdős problems, and the word solved was doing dishonest work. The system had located existing results in the literature for problems still listed as open in an online database, which is a genuine and useful literature-search achievement, not the production of ten original proofs. By 2026 the community had gotten precise about these distinctions. The Erdős problems project's own wiki now tracks full solutions, partial results, failed attempts, rediscovered literature, and Lean-verified proofs separately, and the record holds genuine successes alongside incorrect proofs and results that merely duplicated prior work. One instructive case: a method suggested by GPT-5.4 Pro led to real results on primitive sets and divisibility chains, and the resulting paper names its human collaborators, identifies the model's role, and folds the contribution into a conventional argument rather than presenting the model as a disembodied autonomous discoverer. That is what honest looks like.
The evidentiary ladder above already caught two of the soft variables, the contribution claim, which is autonomy, and the generalization claim. The IMO and Erdős episodes surface two more the ladder does not. One is priority: who produced a result first, who announced first, whether it was formally evaluated or merely posted. Priority is a social and historical claim, not a theorem, and no amount of checking the final equations settles it. The other is difficulty. How long a problem stayed open is not a full measure of how hard it is. Some problems resist sustained assault by many experts; others go unsolved because they are obscure, poorly indexed, or unfashionable, and the Erdős project itself warns that a lack of progress can reflect obscurity rather than difficulty and that its database should not be treated as a benchmark. The Jacobian conjecture was famous and seriously studied, so its resistance was probably real. The general lesson stands anyway. Old and hard are not synonyms.
Notice that all four soft variables point the same direction, toward inflation. Priority framing makes a shared result sound solo. Autonomy framing makes an assisted result sound autonomous. Difficulty framing makes an under-attacked problem sound insurmountable. Generalization makes a narrow win sound like a general one. None of the four ever errs toward modesty, because the announcement is an instrument of persuasion and modesty does not persuade. The proof underneath can be flawless while every dial mounted on top of it is turned, quietly and at once, toward the largest statement the artifact can be stretched to support. A flawless core wrapped in an inflated frame, the one carrying zero discount and the other carrying all of it, is the signature to watch for.
The enterprise lesson is to make the evidence travel #
The right response is not to discount the mathematics. It is to demand that commercial evidence start behaving more like the mathematics.
An enterprise AI claim should ship with an audit package, and it is not a mysterious one: the exact task definition, a representative and independently controlled evaluation set, the full system configuration rather than just the base-model name, the human baseline and the existing-process baseline, the failure categories and exception rates, the number of attempts and the cost per successful outcome, the distribution of performance rather than a single average, the record of human overrides and corrections, and a method a customer or third party can actually rerun. Not every result can be made fully public. Customer data, security requirements, and trade secrets impose real limits. But there is a wide gap between perfect openness and the current habit of publishing a flattering percentage with no reproducible path from the evidence to the conclusion, and almost every vendor is operating deep inside that gap by choice.
The valuable thing about the Jacobian counterexample is not that it happens to be mathematics. It is that the evidence travels with the claim. The commercial market needs far more claims with that structure. Instead of asserting an agent is ninety percent accurate, ship the cases, the scoring rules, and the error taxonomy. Instead of asserting a deployment returned three times its cost, disclose the baseline, the time horizon, the implementation spend, the excluded failures, and the counterfactual. Instead of asserting a model did weeks of work in hours, publish the input, the output, the review time, the correction burden, and the quality bar. The aim is not to abolish judgment. It is to make clear where judgment begins, which is precisely what a self-verifying artifact does and a flattering percentage refuses to.
My own position in this #
A piece arguing for evidentiary discipline should expose its own boundaries, or it fails on its terms and deserves to.
The mathematical core here is independently checkable, and I checked it before I wrote a word of this. The map, the constant determinant, and the three-point collision do not depend on my interpretation, and the conclusion that the conjecture is false in dimensions three and above follows directly. The history of the discovery carries more uncertainty. Alpöge credited Fable, and the credit should be taken seriously, but without the full interaction record no outsider can reconstruct the division of labor with the confidence available for the determinant. Verification arbitrage is more interpretive still. It is my account of how a checkable achievement generates institutional credibility that then colors the reception of unfalsifiable commercial claims. It is an argument, not a theorem.
What the counterexample proves and what it does not #
Take the result at full mathematical weight, because it earns it. A frontier model appears to have participated in discovering a compact, correct counterexample to one of the best-known conjectures in mathematics, an object simple enough for thousands of people to verify independently and surprising enough to redirect a field that spent eighty-seven years pointed the other way. That is a genuine achievement, and it should update your view of what AI-assisted mathematical search can do. The combination of frontier models, expert problem selection, computational tools, and cheap symbolic verification may keep turning up important objects that traditional research processes walked past.
Describe it, though, at the resolution the evidence supports. The map proves the Jacobian conjecture is false in dimension three. It is evidence that an AI-assisted workflow can contribute to novel mathematics. It does not tell us how often the workflow succeeds. It does not establish that the model worked autonomously. It does not establish that the same capability transfers reliably to enterprise operations. And it does not audit anyone's productivity or ROI numbers. The theorem and the sales proposition may share an announcement, but they do not share an evidence base, and that is the final asymmetry.
The most credible AI result of the moment is credible precisely because the audience does not need to trust the lab. Anyone can take the artifact away, run the check, and reach the same conclusion. Where the money actually moves, in the deployments and the payback timelines and the productivity promises, trust is still doing nearly all of the work, because the referee that would make it unnecessary still does not exist. A polynomial map in three variables can settle a mathematical question with no referee at all. It cannot settle the commercial claims made in its shadow.
The theorem needs no discount. Everything inferred from it still does.