The Math Heist: AI Must Be Kingmaker, Not King

Artificial intelligence has entered mathematics in a way that was almost unimaginable a few years ago. It can search enormous spaces of possibilities, generate conjectures, test approaches and work through arguments at a scale no individual mathematician can match.

That is an extraordinary development. It is also exposing a problem that has little to do with computational power.

Mathematics has an economy of recognition.

Unlike many scientific disciplines, pure mathematics often produces no obvious commercial product. A mathematician may spend years proving a theorem whose direct economic value is negligible. The rewards are instead priority, publication, citation, reputation and the knowledge that one’s work has become part of the mathematical record.

AI therefore raises a question that is easy to miss when attention is focused on capability: What happens when machines become capable of producing mathematical results faster than the human system can assign credit for them?

The recent Mathathon controversy, the dispute surrounding OpenAI’s claimed Navier–Stokes solution and concerns raised by mathematicians about AI-assisted research all point toward the same unresolved issue.

The problem is not whether AI should do mathematics. It is how AI should participate in the mathematical ecosystem.

1. The Mathathon Problem

Caltech’s proposed Mathathon was an unusually direct expression of the new model. Teams were to work on open mathematical problems using substantial quantities of AI credits, with more than $2 million in credits reportedly made available through AI-company sponsorship.

The event attracted strong criticism. An open letter signed by more than 700 mathematicians argued that AI companies were turning research mathematics into an advertising opportunity and creating an arms race to announce AI-generated results. OpenAI subsequently withdrew its sponsorship. Business Insider reported on the controversy and withdrawal.

Giving students access to powerful AI systems is not itself a problem. It could produce fascinating mathematics and give young researchers capabilities that previous generations never possessed.

The concern is what happens when the competition to demonstrate AI capability becomes intertwined with the competition to produce mathematics.

Mathematical research traditionally rewards depth, originality and understanding. A technological race naturally rewards speed, visibility and impressive demonstrations. Those incentives are not necessarily the same.

2. Navier–Stokes

The Navier–Stokes episode made the issue impossible to ignore.

OpenAI announced that an internal AI system had produced what it described as a solution to the famous Millennium Prize problem. The system reportedly used around 10,000 AI agents and worked for roughly 88 hours. OpenAI also published a proof and Lean formalization. OpenAI’s announcement is here.

If the mathematics survives the scrutiny expected for a result of this importance, it will be an extraordinary achievement.

Yet the interesting question is not simply whether a machine has solved Navier–Stokes.

A problem such as this is embedded in a century of human mathematics. Thousands of researchers have established partial results, developed techniques, identified obstacles and gradually mapped the territory.

An AI system entering that territory does not arrive in an intellectual vacuum. It arrives carrying the accumulated infrastructure of mathematics.

That is precisely why attribution becomes so important.

3. Who Gets the Credit?

The controversy involving Tristan Buckmaster and Levent Alpöge illustrates the difficulty.

Buckmaster has alleged that OpenAI became aware of their research direction and subsequently pursued a parallel effort. OpenAI disputes the characterization and says it offered arrangements involving publication and recognition. There are also unresolved questions about whether interactions with AI systems could have indirectly influenced model development. These competing accounts should not be treated as established fact. The Financial Times covered the dispute.

But the episode exposes a structural problem even if every disputed allegation were eventually resolved in OpenAI’s favour.

Suppose a mathematician spends months developing an idea and discusses it with an AI system. The conversation contains a conjecture, a failed proof, a useful transformation and a promising direction.

The AI company may control the infrastructure through which that intellectual process takes place. The researcher controls the mathematics.

Those are not the same thing.

This is why AI-assisted research needs much stronger provenance than ordinary software use. We need to know not only which paper was cited, but how the result came into existence. That broader question of who owns and benefits from humanity’s accumulated knowledge is explored in AI, Human Knowledge, and the Missing Social Contract.

4. The Andreas Thom Question

A similar concern emerged around work on non-sofic groups, where mathematician Andreas Thom questioned whether his interactions with ChatGPT could have contributed to OpenAI’s research.

The underlying mathematical work also drew heavily on earlier research, including contributions by Thom and Gábor Kun. Concerns were subsequently raised about attribution and about whether conversations with AI systems might become part of broader model-development processes. The Verge reported on Thom’s concerns.

Again, this is not proof that anyone improperly used private research.

It demonstrates something more fundamental: the researcher increasingly cannot see the entire information pathway through which an AI system operates.

That is uncomfortable in any field. In mathematics, where a single unpublished idea can determine priority, it is particularly consequential.

5. Mathematics Is Not a Factory

There is another dimension to this debate that is even more interesting.

Terence Tao has been warning that mathematical research is not simply a process of selecting a problem and optimizing toward its solution. He describes pure mathematics as one of the unusual serious activities in which curiosity, play, wandering and apparently unproductive exploration can generate important discoveries. Tao’s evolving commentary on AI and mathematics develops this theme in detail.

This matters because AI is exceptionally good at optimization.

A mathematician may spend six months following an apparently useless idea, discover that it fails, learn something unexpected and then formulate an entirely different question.

From an optimization perspective, much of that activity looks wasteful. From the perspective of mathematics, it may be the actual research.

Tao’s warning is therefore deeper than the familiar fear that AI will replace mathematicians. A system might reach the designated answer while eliminating the detours through which human beings acquire understanding and discover the next problem.

I would call that cognitive vertigo: first realizing that the machine can perform the task, then realizing that it can optimize the task, and finally wondering whether some of the things we regarded as inefficiency were actually part of the intellectual value of doing the task.

6. The Problem With the Leaderboard

This is why famous mathematical problems make excellent AI benchmarks but poor definitions of mathematical progress.

There is nothing wrong with asking an AI system to attack Navier–Stokes, the Riemann Hypothesis or another famous problem. Such challenges provide extraordinarily useful tests of reasoning.

The danger comes when the benchmark becomes the culture.

Imagine mathematics increasingly organized around a sequence of spectacular demonstrations: identify a famous unsolved problem, deploy enormous compute, produce a result and announce that AI has defeated another human frontier.

The monetary prize may be $1 million. The real prize is the technological demonstration worth far more to the company.

That is rational from a business perspective. It is not necessarily what mathematics needs.

Tao has also raised a related concern: good open problems may themselves be a scarce resource. If promising problems become targets for enormous automated efforts the moment they are discussed, mathematicians may become less willing to share interesting directions openly. See Tao’s discussion and linked essays.

That would be an extraordinary unintended consequence. AI could make mathematical research more powerful while simultaneously making mathematicians more secretive.

7. From Solving Old Problems to Creating New Ones

The most exciting possibility lies elsewhere.

AI should not merely consume the backlog of problems created by human mathematics. It should help generate mathematical territory that does not yet exist.

A powerful system could examine structures nobody has connected before, suggest conjectures that human mathematicians would never formulate, find unexpected analogies between distant areas or expose patterns buried inside enormous mathematical spaces.

That would change the relationship.

The machine would not simply be competing against mathematicians for existing trophies. It would be helping them discover new mountains to climb.

That is a much more interesting future than an endless AI leaderboard.

8. The Free-Researcher Paradox

There is an additional complication as AI companies place frontier systems directly into the hands of researchers through free or subsidized programmes.

That is potentially wonderful for science.

But a research assistant is no longer merely observing published mathematics. During an interaction it may encounter the researcher’s unfinished thinking: hypotheses, failed approaches, questions and possible discoveries.

This is what I mean by “Spy AI”.

I do not mean that AI companies are necessarily spying on researchers. The phrase describes an information asymmetry.

The AI system may sit inside the process by which unpublished knowledge is being created.

That makes the rules surrounding research conversations extremely important. Researchers should know what is retained, what may enter model-development pipelines, what protections apply and how intellectual contributions can later be traced.

The more useful the AI becomes as a research partner, the more important these questions become.

9. The Shovel Does Not Own the Gold

There is a useful analogy here.

During a gold rush, a company can become extremely successful selling shovels. It does not thereby own every piece of gold discovered with them.

AI is a vastly more sophisticated shovel. It may search, reason, calculate, experiment and propose ideas. Its contribution can be substantial enough that pretending it is merely a calculator would be absurd.

But neither does the opposite follow: because the AI made a major contribution, its owner automatically owns the resulting intellectual discovery.

This is where the familiar Word and Excel analogy remains useful.

Microsoft provides the software. The writer owns the intellectual content of the book. The spreadsheet user does not transfer ownership of every discovery made through Excel simply by using Microsoft’s software.

AI complicates this boundary because it contributes far more actively. That means we need better rules for attribution, not the abandonment of attribution.

10. Recognition Must Become Part of the Protocol

An AI-assisted mathematical paper should eventually make much more visible what happened during its creation.

• Which AI system was used?
• For what tasks?
• Which ideas originated with the researchers?
• Which earlier mathematical work materially enabled the result?
• What was independently verified?
• Were unpublished research materials involved?
• What data and conversations were protected from model development?

These are not bureaucratic questions. They are part of establishing intellectual provenance.

And when a person’s work materially contributes to a result, acknowledgment should be generous.

A bibliography is not always enough.

Someone whose unpublished insight opened the path to a theorem should not disappear into reference number 47 while the AI company receives the headline.

Mathematics has survived for centuries partly because mathematicians care intensely about who discovered what. AI should strengthen that tradition rather than accidentally destroy it.

11. Let Humans Judge the Machine

There is also a cultural question.

AI companies understandably want to demonstrate that their systems are becoming extraordinarily capable. But there is a difference between demonstrating capability and demanding that society accept a declaration of superiority.

If an AI produces a proof that humans independently verify, mathematicians can decide what the achievement means. If the system produces a new conjecture that opens an entire field, the mathematical community can decide how important it is.

The machine can demonstrate.

People should judge.

That distinction matters because intellectual authority cannot simply be transferred to the company operating the most powerful model.

12. Kingmaker, Not King

None of this requires slowing AI down.

Quite the opposite. Mathematics could become one of the greatest beneficiaries of powerful AI.

The machines can explore spaces no human could search, perform tedious derivations, test thousands of possibilities, expose hidden structures and suggest questions that would otherwise remain invisible.

The objective should be to make the mathematical ecosystem larger, not merely faster.

That means preserving the human incentive to explore, maintaining open scientific exchange where possible, protecting unpublished ideas, establishing clear provenance and giving credit to every meaningful contributor—including AI where appropriate. The broader economic question of who should benefit from AI’s use of human knowledge is explored in Universal Basic Equity: Who Owns the AI Future?

The principle is simple.

AI should be kingmaker, not king.

The kingmaker possesses enormous power, but the kingdom is not his merely because he helped someone reach the throne.

AI companies can build the most powerful mathematical tools humanity has ever created. They can make discoveries alongside mathematicians and perhaps eventually make discoveries that humans could not have reached alone.

But the mathematics remains a human intellectual ecosystem.

The shovel can transform the gold rush. It still does not own the gold.