Did AI Solve a $1 Million Problem—or Copy Human Mathematicians?

For nearly two centuries, the Navier–Stokes equations have helped scientists describe how fluids move. They govern everything from water curling around a rock to air flowing over an airplane wing, blood moving through an artery, smoke twisting above a flame, and storms forming in the atmosphere.

There is just one rather awkward problem: mathematicians still cannot fully prove that the equations always behave the way we expect them to.

Now, an artificial-intelligence system may have pushed this mystery toward a solution—and ignited an equally difficult argument about ownership, credit, and the future of scientific discovery.

OpenAI has released a proposed proof concerning the Navier–Stokes existence and smoothness problem, one of the seven Millennium Prize Problems selected by the Clay Mathematics Institute. A valid solution could qualify for a $1 million prize. OpenAI says roughly 10,000 AI agents worked in parallel on the project, exchanging 2.7 million messages and producing approximately 130 billion output tokens.

But the mathematical breakthrough arrived with controversy. Human researchers working on related ideas have questioned how the AI project developed so quickly and whether information about their unpublished work may have influenced it. OpenAI has disputed allegations of improper use.

So did AI independently conquer one of mathematics greatest challenges? Did it assemble clues first developed by humans? Or are we entering an era in which the distinction between those possibilities is becoming impossible to maintain?

1. The $1 Million Math Problem

In 2000, the Clay Mathematics Institute announced seven Millennium Prize Problems, each carrying a $1 million award. They were chosen because they represented some of the most profound unanswered questions in mathematics.

Only one—the Poincaré conjecture—had been resolved before the new Navier–Stokes claim. That history explains why any proposed solution receives intense scrutiny.

A dramatic announcement is not enough. For a result to become accepted mathematics, experts must inspect every definition, inference, and hidden assumption. A proof can be hundreds of pages long and still fail because of a subtle error in a single step.

The Clay Mathematics Institute also has formal conditions for recognizing a solution. Publication and sustained acceptance by the global mathematics community matter. In a September 11 statement, the institute said the problem had “apparently been settled,” while emphasizing that its evaluation process is deliberately unhurried. Until that process is complete, the safest description is not “AI definitively solved Navier–Stokes,” but “AI produced an apparently successful solution that remains under formal review.”

2. What Are The Navier-Stokes Equation

Imagine stirring cream into a cup of coffee. At first, the cream forms long white ribbons. Those ribbons stretch, fold, break apart, and eventually disappear into the darker liquid.

The Navier–Stokes equations are designed to describe motion like this. They connect several physical properties of a fluid, including its velocity, pressure, density, and viscosity. In simple terms, they try to predict where the fluid will move next and how fast it will get there.

Engineers and scientists use versions of these equations to model:

airflow around aircraft and automobiles;
ocean currents and weather systems;
fuel moving through engines;
blood circulating through the body;
water traveling through pipes;
smoke, flames, and industrial chemicals;
the formation of turbulent eddies and vortices.

Computers can approximate the equations extremely well in many practical situations. That is how engineers simulate airflow before building an aircraft or estimate how a storm might develop.

3. How Thousands Of AI Agents Approached The Problem

The system reportedly did not operate like one person having one brilliant flash of insight. Instead, many AI agents worked in parallel.

One agent could attempt a construction. Another could search for a contradiction. Others could expand technical lemmas, test alternative strategies, compare the argument against earlier literature, or act as skeptical reviewers. Promising branches could be developed while failed branches were discarded.

This resembles a vast research group that never sleeps, reads at machine speed, and can pursue thousands of mathematical paths simultaneously.

The scale is precisely what makes the result so fascinating—and unsettling. Human mathematicians often spend years building the intuition required to recognize a fruitful direction. An AI system can use massive computation to explore more directions than any individual could examine in a lifetime.

But computation alone does not guarantee truth. Language models can generate confident errors, reproduce a hidden assumption across several agents, or create an argument that looks rigorous until a specialist tests its weakest point. Ten thousand agents repeating the same mistake do not transform that mistake into a theorem.

4. Who Should Receive Recognition Credit

An AI system cannot meaningfully spend $1 million, sign an authorship agreement, accept ethical responsibility, or defend its conduct. Behind it stand model developers, mathematicians, computing infrastructure, data providers, and innumerable authors whose work helped create its knowledge base.

Possible answers include awarding recognition to:

the researchers who designed and directed the AI experiment;
the mathematicians who supplied the decisive ideas;
the organization that financed the computation;
all significant human contributors;
no one, if the result cannot be assigned to a responsible author.

Each option creates problems. Rewarding only the AI’s operator may erase intellectual contributions from others. Dividing credit among everyone who influenced the model may be impossible. Denying recognition to AI-assisted work could also discourage a method capable of producing valuable discoveries.

The real solution may require a new category of scientific authorship—one that separates conceptual contribution, computational execution, verification, and responsibility.

5. Did AI Actually Understand The Proof

One view is that understanding requires intuition—the ability to see why an argument works, connect it to physical reality, and explain its deeper meaning. On this view, producing a formally valid sequence of steps is not the same as grasping the mathematics.

The opposing view is more pragmatic. If a system can generate a genuinely new proof, defend each step, correct objections, and use the result to solve new problems, refusing to call that understanding may become increasingly difficult.

Humans also rely on tools. No one says an astronomer failed to make a discovery because a telescope collected the light, or that a biologist deserves no credit because software analyzed the genome. Perhaps AI is simply the next scientific instrument.

Yet ordinary instruments do not usually propose theories, write proofs, debate reviewers, or potentially compete with their operators for recognition. AI is not just extending human perception; it is beginning to occupy parts of the reasoning process itself.