AI is solving maths faster than anyone can check it. Your company is next.
Mathematicians are already retracting machine-made papers within a day, and most businesses have no comparable way to test what AI produces.

Sign up for my Substack daily AI newsletter https://johanosteyn.substack.com
Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/
Mathematics used to be the one field where a result was either right or wrong, and someone could prove which. Artificial intelligence is now producing mathematics faster than mathematicians can check it. The lesson for boards is not about equations. It is about what happens to sign-off when machines produce work faster than people can understand it.
CONTEXT AND BACKGROUND
The progress is real. An AI system made progress on a question Paul Erdős posed in 1946, overturning a long-standing expert prediction, and OpenAI announced that its agents had produced a solution to a version of the Navier-Stokes problem in 88 hours, a claim still awaiting independent verification. Northeastern University's reporting also notes that some AI solutions have been retracted when they failed to hold up, and that AI often cannot explain how it reached a proof.
The strain is showing. On 6 October, OpenAI posted 722 preprints claiming progress on 372 problems, and roughly half of the results were released unconfirmed. The next day it withdrew three manuscripts after a sign error broke the argument in one and the construction behind two others. MIT's Andrew Sutherland called the withdrawal responsible, but said it "will take a lot more than that to earn back the trust they have lost".
Even the checking tools can fail. In July, an AI-assisted "disproof" of the Collatz conjecture passed Lean, the formal proof checker mathematicians rely on, because of a bug in Lean's own core. A second independent checker also had a flaw. Both were fixed quickly, and the conjecture remains open.
INSIGHT AND ANALYSIS
The deeper problem is understanding. Terence Tao, one of the world’s leading mathematicians, warned at the International Congress of Mathematicians: “We can have these 100,000-line proofs that we have to verify, but no one understands them.” He added that “It is not enough to generate and verify the proofs.” At a Stanford symposium, he proposed a simple test of whether a result has been digested: can someone give a talk about it and take questions? Leonardo de Moura, who created Lean, said that checking whether a formal statement matches what a paper intends is “still going to be a human activity.”He added that “It is not enough to generate and verify the proofs”.
I have previously written about this in a different form, arguing that a model passing a test does not prove that it can be trusted. Mathematics now shows the same gap in its purest form.
The fair objections deserve a hearing. Verification tools are improving quickly, and the Lean bug was fixed within hours of being reported. Human proofs have always contained errors, and retraction is how science corrects itself. Most business output is judged by results rather than logic. All true. Yet mathematics has the best checking machinery of any field, and it is still struggling. Most companies have nothing comparable for AI-generated analysis, code, contracts or forecasts. A spreadsheet does not need a formal proof, but the weakness is the same: output nobody has checked and nobody can explain.
IMPLICATIONS
In South Africa, the people best equipped to check machine reasoning are in short supply. In the 2025 National Senior Certificate, 162,947 of 764,014 full-time candidates passed mathematics, or 21.33%. In March 2025, the Minister of Basic Education reported that 464 public schools do not offer mathematics at all, citing among other reasons a shortage of qualified teachers. Those numbers feed the pipeline of actuaries, analysts, engineers and data scientists who would audit automated systems. A country that produces too few people able to follow a chain of reasoning will be the one most tempted to take AI's answers on trust.
Boards should ask three questions. First, what share of AI-generated work is actually checked before it is used, and by whom? Second, can someone in the organisation explain the result and defend it under questioning, not merely confirm that it ran? Third, what does an error cost when it is discovered the day after release rather than the day before?
CLOSING TAKEAWAY
AI's progress in mathematics is remarkable, and it will spread to every field that runs on reasoning. The constraint will not be producing answers. It will be checking and understanding them. Companies that keep people able to do that will govern their AI. The rest will be signing off on work nobody in the room can explain.
Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence's impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan's work at https://www.aiforbusiness.net




Comments