On September 8, OpenAI announced that an internal AI system had produced a solution to the Navier-Stokes Millennium Prize Problem, one of the most famous unsolved problems in mathematics.
The scale alone is remarkable. Roughly 10,000 agents worked on the problem. The successful effort took about 88 hours, generated 2.7 million agent messages and around 130 billion output tokens, followed by another 17 hours of formalization and verification in Lean.
Three days later, the Clay Mathematics Institute said the problem had “apparently been settled,” while making clear that evaluating the result and assigning credit would be a deliberately careful process. Since then, the question of credit has become part of the story itself. Twenty-five Fields Medal winners have raised concerns about attribution and plagiarism as AI systems absorb, combine, and extend mathematical work.
That may prove more consequential than the proof.
For centuries, accomplishing something extraordinarily difficult was evidence that you have extraordinary expertise. Solve the theorem. Discover the molecule. Develop the theory. Earn the recognition.
AI is beginning to separate achievement from mastery.
A person can now direct a system capable of producing a breakthrough they could never reproduce themselves and may not fully understand.
When intelligence becomes abundant, accomplishment stops being proof of expertise.
Science Was Built Around Scarce Intelligence
The historical comparison I find most useful is the printing press.
The printing press dramatically reduced the cost of distributing knowledge. The internet reduced the cost of accessing it.
AI could reduce the cost of producing new knowledge.
That changes the economics of science because much of our scientific system was built around the assumption that exceptional intellectual capability is scarce.
Universities compete for brilliant researchers. Grant programs decide which researchers deserve funding. Journals compete to publish important discoveries. Prizes identify exceptional individuals. Academic careers are built around a record of difficult intellectual accomplishments.
Now imagine thousands or millions of AI research agents exploring hypotheses simultaneously.
We are already seeing the early version of that. The Navier-Stokes effort required coordination among thousands of agents, which connects directly to the challenge I explored in Single Agent vs. Multi Agent AI. At that scale, producing intelligence becomes an infrastructure and orchestration problem.
Science could move from a world where answers are scarce to one where good questions are scarce.
That changes the scientist’s highest value contribution.
Which problem deserves solving? Which assumption should we challenge? Which result matters? What should we test next? Can we trust the answer?
When answers are cheap, questions are priceless.
The scientist remains essential. Being the person who simply has the answer becomes far less distinctive.
A Breakthrough May Need an Audit Trail Instead of an Author
The obvious answers to the ownership question all have problems.
The company that built the AI may own the system without having conceived the scientific insight. A researcher may have supplied the prompt without performing most of the intellectual work. Giving ownership or authorship to the AI creates an entirely different accountability problem.
My view is that credit should follow intellectual contribution.
Science needs a richer concept than traditional authorship. I would think of it as an intellectual provenance chain.
Who identified the problem?
Who contributed original insight or unique data?
What did the AI generate?
Who validated the result?
Who recognized why it mattered?
That record tells us far more than a name at the top of a paper.
It also protects science from a very predictable problem. Someone should not be able to enter a sophisticated question into an AI system, receive a major discovery and claim personal genius for everything that follows.
Prompting a discovery is not necessarily the same as making one.
We are entering an era where the breakthrough may have an audit trail instead of a single author.
Accountability matters here too. As I argued in Who’s Responsible When an AI Agent Makes a Bad Decision?, delegating cognition to a machine does not make responsibility disappear. Science will need to know who stands behind a result when the system that generated it cannot.
What Happens When Discovery Outruns Peer Review?
There is another institutional problem hiding behind the excitement.
Scientific production may soon scale much faster than scientific validation.
If thousands of research agents can generate proofs, hypotheses, candidate drugs, and experimental directions around the clock, human experts cannot review every result at the same speed.
Peer review then becomes a bottleneck.
The highest value work shifts toward verification, replication, provenance, and deciding which machine-generated results deserve human attention.
Formal mathematics has an advantage because systems such as Lean can verify whether a proof follows logically. Most scientific disciplines do not have an equivalent. A proposed treatment still needs experiments. A new material still has to be manufactured and tested. A biological hypothesis still has to survive contact with the physical world.
This means AI can make one part of science dramatically faster while increasing pressure on everything downstream.
Institutions will have to redesign review around that new volume. Otherwise we create an abundance of potential discoveries and a shortage of trusted ones.
Can we trust the answer may become one of the most valuable scientific questions of all.
Funding Will Follow the New Scarcity
Today we largely fund researchers and institutions and ask them to produce discoveries.
That model makes sense when intelligence itself is one of the scarce resources.
If AI expands research capacity dramatically, funding starts moving toward different bottlenecks: important problems, proprietary data, compute, experimental infrastructure, and validation.
The grant question changes.
Instead of asking which scientist deserves ten million dollars, we may increasingly ask which unanswered question deserves ten million dollars of intelligence.
That could democratize scientific discovery by giving smaller teams access to capabilities once concentrated inside elite institutions.
It could also concentrate scientific power even further.
Whoever controls the largest compute resources, the best proprietary datasets, the laboratories required for validation, and the infrastructure needed to operate large research systems gains enormous leverage. This is closely related to the infrastructure argument I made in The Next Superpower Won’t Be a Country.
AI will not eliminate scarcity from science. It will move scarcity from intelligence to questions, data, compute, and judgment.
We can already see a version of this in drug discovery. As I explained in The Next Billion Dollar Drug Was Designed by an Algorithm, generating a promising candidate is only part of the process. Validation, experimentation, and deciding what deserves further investment remain essential.
Academic Careers Have a Different Problem
There is a human issue underneath all of this.
Scientists spend years building an identity around mastering difficult things. So do physicians, engineers, professors, programmers, and lawyers.
The difficulty matters. It becomes part of the credential and part of the person’s professional identity.
What happens when someone using AI can produce an output that once required twenty years of expertise?
For most of history, doing something extraordinary made you extraordinary.
AI is breaking that relationship.
Academic careers will have to adapt because publication volume and technical output become weaker signals of individual ability when machines can generate more of both.
Prestige may gradually move toward different evidence of value: identifying important problems, designing good experiments, exercising judgment, validating results, connecting discoveries across fields, and taking responsibility for what gets pursued.
That transition will be uncomfortable because our institutions have spent centuries rewarding answers.
They may soon need to reward better questions.
What Does Discovery Mean When Nobody Understands It?
There is an even stranger possibility.
Imagine an AI discovers a new material, treatment or mathematical relationship that repeatedly works, yet no individual human can fully explain why.
What exactly have we discovered?
Historically, discovery implied some degree of understanding by the discoverer. AI breaks that connection.
Can humanity know something that no individual human understands?
We already worry about people outsourcing parts of their thinking to machines, which is the problem I explored in We’re Raising a Generation That Outsources Thinking. At the scientific level, the stakes become much larger. Entire fields could eventually contain useful knowledge that humans can validate without being able to reproduce the reasoning that generated it.
AI creates the first human generation capable of achieving things they do not actually understand.
That should make us curious rather than fatalistic.
Human value does not disappear when machines become better at producing answers. It moves toward curiosity, judgment, meaning, validation, responsibility, and deciding what deserves to exist.
Scientific prestige may eventually shift from “I found the answer” toward “I asked the question nobody else thought to ask.”
That could produce an extraordinary period of discovery.
It will also force us to redesign institutions built around a simple assumption that has held for centuries: the mind responsible for a scientific breakthrough belongs to a human being.
That assumption no longer holds.
Now science has to decide what replaces it.
