AGI Made Clear › Research and verification
What happens when AI produces research faster than humans can verify it?
OpenAI says an internal model has resolved more than 100 long-standing mathematical problems in under a month. The claim still needs scrutiny, and no public catalogue has been published. But it raises a larger question: what happens if producing new research becomes faster than checking it?
For most of scientific history, discovering something new has been the difficult part.
A mathematician might spend years trying to prove a theorem. A scientist might spend years developing and testing a hypothesis. Then other experts would examine the work, challenge it, reproduce it, or find the mistake.
That system rests on an assumption so basic it is rarely stated: new knowledge is hard to produce.
AI may be starting to challenge it.
On 21 September 2026, OpenAI announced that an internal model it began training on 28 August had, in its words, "resolved more than 100 long-standing open problems across most areas of mathematics".
OpenAI says the pace surprised its own mathematicians. Enough that it has set up an advisory group of mathematicians to work out how AI-generated results should be evaluated and released.
The obvious headline is that AI solved a hundred maths problems.
That is probably not the important part.
First, what the claim actually is
We should be careful with the word solved.
OpenAI says its model has resolved more than 100 long-standing open problems. That is a statement by the company about its own internal system.
The announcement does not come with a public catalogue of the results. So there is not, at the time of writing, a list that mathematicians outside OpenAI can work through and check.
That matters more than usual here, because verification is precisely what this story is about. We should not treat the number as though the mathematical community has examined and accepted a hundred new proofs. It has not had the chance.
What we can say is narrower and still significant: OpenAI believes its internal systems are producing potentially important mathematical results at a rate high enough that it thinks new processes are needed to handle them.
Whether every individual result survives scrutiny is almost secondary to the institutional problem that creates. Even if only a fraction hold up, the volume changes how mathematical research works.
The Navier-Stokes result gives a sense of the scale
This did not come out of nowhere.
On 8 September 2026, OpenAI published what it describes as an AI-generated proof, with a Lean formalisation, that an initially smooth fluid at rest can develop a singularity in finite time. Navier-Stokes existence and smoothness is one of the seven Millennium Prize Problems.
The system was not a chatbot answering a question. OpenAI says it used coordinating groups of agents with access to tools, and that the group which produced the Navier-Stokes result "involved on the order of 10,000 concurrent agents".
Two details are worth holding on to.
OpenAI published a Lean formalisation alongside the written proof, which means the logical structure can be mechanically checked rather than taken on trust. That is a meaningful difference from a claim in a press release.
And OpenAI explicitly says it does not intend to claim the Millennium Prize for the result, describing the announcement as "a snapshot in time". A company that believed the matter settled would not phrase it that way.
AI changes the economics of trying things
Human mathematicians have limited time. A researcher might see ten possible approaches to a hard problem and have time to pursue two.
Ten thousand agents do not face that constraint in the same way.
Most attempts fail. Many reach dead ends. Some produce nonsense. But if one finds a valid path, the economics of mathematical exploration have changed.
This may be one of AI's most important effects on research, and it is easy to miss. The system does not need to reason like the greatest mathematician alive. It may simply be able to attempt vastly more lines of reasoning than any human group could.
That is a different kind of advantage: scale of exploration rather than depth of insight.
Generating a proof and trusting a proof are different problems
Imagine a system produces one potentially important proof a month. Mathematicians can examine it properly.
Now imagine ten. Then a hundred. Then a thousand.
At some point the bottleneck moves. The hard question stops being whether we can generate a possible solution, and becomes whether we can establish which solutions are correct.
Mathematics is unusually well placed for this, because some proofs can be formally verified. Tools such as Lean can check whether an argument follows from its assumptions.
But formal verification does not remove every problem. Researchers still need to know whether the problem was formulated correctly, whether the result is genuinely new, whether equivalent work already exists, whether the assumptions are meaningful, how important it is, how it connects to what is already known, and who deserves credit.
Some of those are mathematical questions. The rest are questions about institutions.
The advisory group, and the line it will not cross
OpenAI's response to this is itself revealing.
The company has set up an Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study. It has ten founding members, among them Timothy Gowers, Martin Hairer, Edward Witten, Camillo De Lellis, Ulrike Tillmann, Ravi Vakil, Nikhil Srivastava, Francois Charles and Melanie Matchett Wood.
OpenAI says the members will not be paid by the company, that the group operates independently, and that it is free to offer unsolicited advice and make its recommendations public.
It will advise on reviewing and communicating emerging results, assessing their significance, coordinating how they are disseminated, academic standards, and how AI tools can support mathematical research.
And then there is the limit, in OpenAI's own words: "The group will not be responsible for advising us on how to pace our internal progress on mathematics."
Read that alongside the rest and the shape of the arrangement becomes clear. The AI system may keep producing results at increasing speed. The human institution exists to work out how those results should enter mathematics. But the institution has no say over the speed.
Generation and verification are being deliberately placed on separate tracks, with independent oversight on only one of them.
Now imagine this across science
Mathematics may simply be where this becomes visible first, because mathematical proof is unusually checkable.
Picture similar systems in physics, chemistry, materials science, biology, medicine, engineering and computer science.
Instead of researchers asking AI for help with individual tasks, laboratories could run thousands of research agents continuously. One group proposes hypotheses. Another searches the literature. Another designs experiments. Another analyses results. Another attacks the conclusions. Another writes them up.
Research starts to look less like a scientist with an assistant and more like an automated research organisation.
The volume of candidate discoveries could rise enormously. Human attention will not.
And now it is happening in biology
Mathematics is not the only field where this is starting to matter, and the second example arrived two days after the first.
Anthropic has reported the first results from its new life sciences laboratory, where Claude agents search large biological datasets, generate hypotheses, and work alongside human scientists who can test those ideas experimentally.
One of those searches produced something.
Anthropic says around 950 Claude agents spent roughly 21 hours searching a DNA sequence database, consuming 210 million tokens. They gathered more than 200,000 reverse transcriptases, a class of enzyme involved in copying genetic information, identified about 3,500 candidate systems, and narrowed that to twenty worth a closer look.
Near one of the enzymes was a long array of evenly spaced DNA repeats, a pattern reminiscent of CRISPR.
That led to what Anthropic calls ART, array-associated reverse transcriptases, a previously uncharacterised system found mainly in bacteriophages, the viruses that infect bacteria.
Anthropic is careful about what this does and does not mean, and the caveats deserve as much attention as the finding.
It says it does not yet know the system's function, and that work to understand it is ongoing. It says Claude "appears to be the first to notice the system's defining features", which is a narrower claim than first discovery. And it frames the decision to publish as sharing findings early to show what the approach can do, rather than as establishing importance.
So this is not "Claude discovered the next CRISPR". Nobody knows that, including Anthropic.
What is interesting is the shape of the process.
Anthropic says human involvement was "limited to the initial prompt and the lab work", and that the agents used their own judgement to identify which candidates were worth pursuing. Hundreds of agents explored an enormous search space at once. One noticed an unusual pattern. Others investigated it. Human scientists then did the physical experiments needed to find out whether the idea was real.
That is the mathematics problem again, in a different discipline.
And biology adds a complication mathematics does not have. A proof can ultimately be checked against formal rules. A biological hypothesis has to meet the physical world. Cells have to behave as predicted. Proteins have to do what was expected. Experiments have to replicate. Other laboratories have to get the same result.
There is no formal verifier for that. There is laboratory time, and there is a fixed amount of it.
So scale this up. Not 950 agents for 21 hours. Ten thousand. A hundred thousand. Running continuously across biology, chemistry, materials science, medicine and mathematics.
The bottleneck stops being the generation of plausible ideas. It becomes deciding which of an enormous number of AI-generated ideas deserve scarce human laboratory time.
That changes the economics of research in a way that is easy to underestimate. For most of scientific history, good hypotheses have been the scarce thing. We may be moving towards a world where hypotheses are abundant, and judgement, experimental capacity and verification are what run short.
The AI did not replace the scientists here. It expanded, by a large factor, the number of things the scientists could look at.
For now that distinction holds. The direction of travel is the thing to watch.
Scientific publishing was not built for this
Science has institutions for handling new knowledge: peer review, journals, conferences, replication, citation, professional societies.
All of them evolved around human research speed. A paper takes months to review. An important claim takes years to establish. Careers are built on deep expertise in narrow areas.
What happens if AI generates thousands of credible-looking research outputs in that same period?
We could end up somewhere odd: an abundance of potential knowledge and a shortage of trusted knowledge.
The scarce resource becomes verification.
And verification will probably automate too
There is an obvious response. Use AI to check AI.
If one system produces a proof, another attacks it. If one agent proposes a hypothesis, another searches for contradicting evidence. If one model writes code, another tests it. Where the argument can be formalised, a proof assistant checks the logic.
This may become essential rather than optional.
But it raises the question the whole story keeps circling. If AI generates the research and AI verifies the research, where exactly does human judgement sit?
Probably in layers: AI generates, AI challenges, formal tools verify where verification is possible, other systems reproduce, and humans evaluate significance and consequences.
That could increase throughput enormously. It would also change what a researcher is for.
The scarce skill may move
Today a researcher spends enormous effort obtaining a result.
In an AI-accelerated environment the hard part may become choosing: which questions deserve attention, which results deserve verification, which discoveries matter, which experiments are worth running, which risks are acceptable, and which knowledge should be acted on.
Human judgement does not disappear. Its position in the process moves.
The scarce skill shifts from "can you derive the answer" towards "do you understand what this answer means".
Credit becomes genuinely difficult
Suppose a mathematician chooses a problem. Agents explore fifty thousand approaches. Another model identifies the promising ones. A proof system verifies the result. A human checks its significance.
Who discovered it?
The researcher who chose the problem? The team that built the system? The company that trained the model? The mathematicians who verified it? All of them?
This is not merely philosophical. Academic careers, research funding, prizes and reputations all depend on attribution. The institutions around research may have to change alongside the technology.
And then there is rediscovery
A system may produce something it believes is new.
But mathematics contains an enormous amount of existing work. Some results sit in obscure papers. Some appear in slightly different form. Some are known to specialists and barely documented.
As AI-generated research scales, establishing whether a result is actually novel gets harder, not easier.
AI could help with that too, by searching and comparing the literature. Which lands us back in the same pattern: AI generates more information, and we need more AI to manage the information AI generated.
Why this connects to the AGI question
Scientific research has long been treated as an important test of advanced AI.
Most benchmarks measure whether a system can answer questions humans already know how to answer. Research is different: the answer may not exist yet.
A system that repeatedly contributes genuinely new mathematical knowledge is doing something more significant than recall or exam performance.
That does not make it AGI. There is no accepted threshold at which solving some number of problems makes a system generally intelligent, and anyone claiming otherwise is inventing a line.
But discovery is one of the capability areas worth watching closely, particularly if it starts appearing across several disciplines rather than one.
And to AI building AI
AI research is itself a scientific and engineering activity.
If capable models accelerate mathematics, coding and science, they may also accelerate the research required to build better AI.
That is the feedback loop that keeps appearing in discussions of AGI and recursive self-improvement: AI accelerates research, some of that research improves AI, better AI accelerates research further. Taken to its limit, that is the argument for an intelligence explosion.
No system needs to redesign itself overnight for this to matter. Partial automation of research is enough to change the rate.
OpenAI appears aware of the connection. It has published material on how coding agents are changing its own researchers' workflows, and argued that the public needs visibility into how frontier AI is accelerating research, because that may become relevant to decisions about how fast frontier systems should advance.
The evidence still matters
This is where it is worth resisting the pull of the story.
"AI resolved more than a hundred open mathematical problems" is an extraordinary claim, and extraordinary claims deserve verification rather than repetition.
Some results may prove important. Some may be rediscoveries. Some may resolve narrow questions. Some may contain errors. Some may need substantial human correction. Some may survive scrutiny and genuinely add to mathematics.
Until the work is examined, nobody knows the distribution, including OpenAI.
The notable thing is that the verification process has itself become part of the story.
What to watch next
How many results are published. A substantial public catalogue would let mathematicians outside OpenAI assess the claim. Its absence is currently the single biggest gap.
How many survive independent scrutiny. This matters far more than the number generated.
How hard the problems were. "Open problem" covers an enormous range. Ten genuinely deep results would matter more than hundreds of narrow ones.
How much human involvement was required. There is a large difference between a system developing a proof and a system contributing to a process directed throughout by expert mathematicians.
Whether results can be formally verified. The Lean formalisation of the Navier-Stokes work suggests one route to scaling verification.
Whether this happens outside mathematics. If comparable acceleration appears in physics, biology, chemistry and engineering, the implications grow considerably.
Whether AI research itself accelerates. This may be the most consequential measurement of all.
The bottleneck may be moving
For centuries, civilisation has had a shortage of answers. More questions than people capable of answering them.
AI could begin to change that.
If it does, the problem does not disappear. It moves.
From how do we discover something new, to how do we know which discoveries are true, and then to which of those truths actually matter.
That would be a profound change in how knowledge gets made.
And OpenAI's claim about a hundred results may end up mattering less for the number than for the problem it creates: what happens when discovery moves faster than the institutions built to understand it.
What this article does not establish
It does not establish that more than 100 open problems have been solved. That is OpenAI's characterisation of its own internal system's output, published without a catalogue of results, and it has not been independently checked.
It does not establish that the Navier-Stokes result is correct. OpenAI published a proof and a Lean formalisation, says it does not intend to claim the Millennium Prize, and describes the announcement as a snapshot rather than a conclusion.
It does not establish that these systems are doing mathematics in the way a mathematician does. Scale of search and depth of insight are different things, and the announcements do not distinguish between them.
It does not establish that verification is currently a bottleneck. It establishes that OpenAI expects it to become one, which is why the advisory group exists.
It does not establish that ART is significant, or novel in the strongest sense. Anthropic says its function is unknown, that work to understand it is ongoing, and that Claude "appears to be the first to notice the system's defining features" rather than claiming outright discovery. Human scientists carried out the physical experiments.
What it does establish is that two frontier laboratories, in two different disciplines, are now generating candidate results faster than the systems for checking them were built to handle. One of them has created an institution to manage that output, and has explicitly placed the speed of the system outside that institution's remit.