Mathematics has traditionally had a rather pleasant problem. Proofs are difficult to produce.
You can spend months staring at a problem, fill several notebooks with increasingly desperate symbols, drink an irresponsible amount of coffee, and eventually discover that the thing you were trying to prove was false. This scarcity created the entire machinery of modern mathematics.
A researcher finds a result. Someone checks it. The author explains it. A journal publishes it. Other mathematicians read it.
Eventually, if the result matters enough, it enters textbooks and becomes part of mathematical culture. It’s slow. But it works.
Terence Tao thinks AI may be about to break this system. Not because AI mathematics is necessarily wrong. Because it may become too productive.
Mathematics Has Had a Crisis Before
At the International Congress of Mathematicians 2026 in Philadelphia, Tao gave a lecture titled Mathematics in the Age of AI. His historical comparison was striking.
Mathematics, he argued, may be entering a period comparable to the foundational crisis of 1900–1930. Back then, mathematics discovered uncomfortable things about itself.
Russell’s paradox attacked naive set theory. Gödel demonstrated fundamental limits to formal systems. Mathematicians had to reconsider some of the assumptions underneath their discipline.
The current crisis is different. AI isn’t forcing mathematics to reconsider logic. It’s forcing mathematicians to reconsider what mathematical work is valuable. That’s potentially an even stranger problem.
First, Ignore the Marketing
Tao is appropriately cautious about claims that AI is already transforming mathematical research. And mathematicians are perhaps the ideal population for this particular attitude. Show them a spectacular demo and they’ll politely ask for the proof.
Many public claims about frontier models come from uncontrolled experiments. Successful examples are disproportionately reported. Compute budgets aren’t always disclosed. Human intervention varies. Companies have obvious commercial incentives to announce impressive results.
So Tao points toward something much more interesting: First Proof. The project was explicitly designed to evaluate AI on previously unpublished, research-level mathematics rather than familiar benchmark questions. Its second formal benchmark selected ten problems from actual mathematical research and tested multiple AI systems under controlled conditions, with solutions evaluated by expert referees.
The results are already difficult to dismiss as a party trick. The strongest participating agent system ultimately solved six of ten research-level problems, according to ETH Zurich’s account of the challenge. Not Olympiad exercises. Research mathematics. That distinction matters enormously.
The Five Stages of Mathematics
The most useful part of Tao’s argument isn’t actually about model capability. It’s his decomposition of mathematical progress. A result doesn’t become mathematics merely because someone produces a proof.
There are roughly five stages:
1. Discovery. Find the proof.
2. Verification. Determine whether the proof is actually correct.
3. Exposition. Turn it into something another mathematician can understand.
4. Publication and acceptance. Get the community to inspect, debate, reference, and trust it.
5. Canonization. Integrate the result into the accumulated structure of mathematics: surveys, techniques, textbooks, courses, and eventually the mental models of future mathematicians.
AI is becoming surprisingly good at stage one. It’s making serious progress on stage two, especially when formal verification systems such as Lean enter the loop.
Stages three through five? Those still move at approximately human speed. And humans have not recently received a major hardware upgrade.
Welcome to the Proof Supply Chain Crisis
This creates a wonderful inversion. For centuries, the bottleneck was producing mathematics.
Now imagine that producing candidate proofs becomes cheap. Very cheap. An agent runs overnight. Another runs for a week.
Thousands of agents attack thousands of problems. Suddenly the scarce resource isn’t proof generation. It’s attention.
A proof waits to be checked. A checked proof waits for someone to understand it. An understood proof waits for someone to explain it. An explained result waits for reviewers. A published result waits for the community to decide whether anyone should care.
Mathematics moves from scarcity to abundance. And abundance sounds wonderful until you remember what happened to email.
The Verification Queue Is Already Appearing
This isn’t entirely hypothetical. AI systems are already being deployed against open mathematical problems. One recent project used AI-driven formal proof search on hundreds of open Erdős problems, resolving several while producing machine-checkable formal proofs. But there’s an increasingly awkward phenomenon around AI-generated mathematics: people submit candidate solutions they cannot personally verify. Think about how strange that is.
Historically, if you claimed to have proved a theorem, the minimum expectation was that you understood your proof.
Now we can have: “Here is my proof.” “Can you explain step 17?” “No idea. Claude wrote it.”
Mathematics has not really developed social machinery for this. And it probably needs to.
The Most Interesting Possibility Is Also the Weirdest
Push this trend far enough and we arrive at something genuinely uncomfortable. Suppose an AI produces a proof of an important theorem. Suppose a formal verifier confirms every logical step.
The proof is correct. But it’s enormous. It uses unfamiliar constructions. It jumps between areas of mathematics in ways humans don’t naturally think about. No mathematician can develop a useful conceptual understanding of why it works. What exactly have we obtained?
Technically: knowledge.
Culturally: perhaps not.
This distinction may become extremely important. A verified theorem sitting inside a database isn’t necessarily the same thing as mathematics that humans possess.
Proof Was Never the Whole Product
This is where Tao’s argument becomes much broader than mathematics.
AI forces us to separate production from understanding. We’ve spent decades rewarding the visible output. Solved the problem. Wrote the code. Produced the report. Found the theorem. But when production becomes cheap, everything around production becomes more valuable.
Judgment. Verification. Explanation. Context. Selection. Knowing which result matters. Knowing how it connects to everything else.
Turning ten thousand pages of correct mathematics into three ideas another mathematician can actually use. The prestige hierarchy may therefore have to change.
The mathematician who “solved it first” may become less important. The mathematician who explains what the solution means may become considerably more important.
Tao’s Rules for the AI Era
One recommendation is wonderfully straightforward:
Stop pretending you didn’t use AI.
If AI materially contributed to mathematical work, disclose it.
Hidden use creates worse incentives than acknowledged use because researchers become afraid of professional stigma while everyone privately suspects everyone else anyway.
The second principle is more demanding. If authors cannot explain their own result coherently, at an expert level, with appropriate context and references, perhaps it isn’t ready for publication. That’s a high bar. It should be. Because once generating mathematical material becomes cheap, filtering mathematical material becomes the job.
Mathematics May Be the Preview
I suspect this is why Tao’s lecture matters outside mathematics. Math is simply encountering the problem early because correctness is unusually explicit there. Software engineering is heading toward exactly the same situation. So is scientific research. Legal analysis. Financial modeling. Academic publishing.
Anywhere AI can produce intellectual artifacts faster than humans can evaluate them, the same bottleneck appears:
generation → verification → explanation → acceptance → knowledge.
AI accelerates the first box enormously. It doesn’t automatically accelerate the remaining four. And that changes what humans are for.
Final Thoughts
The popular version of the AI-and-mathematics story is: “AI can solve hard math now.” That’s interesting. But Tao’s version is much more important.
What happens when solving mathematical problems is no longer scarce? Mathematics has spent centuries building institutions around a world in which proofs are difficult to obtain. We may now be entering one where proofs are abundant and human comprehension is scarce. That isn’t the end of mathematics.
It may actually produce more mathematics than humanity has ever seen. But it forces the profession to reconsider what deserves status, attention, publication, and ultimately a place inside human knowledge.
The great mathematical bottleneck may no longer be finding the proof. It may be finding someone who has time to understand it. And, for once, “too much mathematics” may become an entirely serious problem.