A beginning for mathematics

[mklabgr]

Summary

In “A beginning for mathematics,” Daniel Litt, professor at the University of Toronto, argues that the rapid progress of AI in mathematics should not be understood primarily as “the end of mathematics,” but potentially as the beginning of an extraordinary new era. He assumes that AI systems will soon become robustly superhuman across most mathematical tasks and argues that the key question is therefore not how to stop AI from producing mathematics, but how humans can preserve and deepen mathematical understanding when producing proofs and solving problems becomes increasingly automated. 

1. Mathematics is more than proving theorems
Litt begins by questioning the way mathematics has traditionally measured success. The mathematical community has largely operationalized its work through theorems, proofs, papers, and solutions to open problems, but these are only imperfect proxies for the deeper goals of mathematics. A machine could theoretically enumerate axioms, propositions, and proofs without possessing anything resembling human understanding. Even if AI eventually produces excellent mathematics, Litt argues, the production of mathematical truth and the human understanding of that truth are different things. 
He identifies two broad goals for the profession:

1. Producing and understanding high-quality mathematics.
   
2. Producing high-quality mathematicians.
   

The second goal becomes particularly important in an AI-dominated environment.

2. AI may make mathematical understanding more important
Litt's central optimistic thesis is that an explosion of AI-generated mathematics could actually create more need for mathematicians, not less. If AI can generate thousands or millions of interesting results, humans will still need to decide what is meaningful, understand the results, connect them to existing ideas, explain them, and determine what questions deserve further investigation. 
He therefore wants mathematics to place greater emphasis on activities such as:

• mathematical conversation;
  
• seminars;
  
• asking good questions;
  
• developing intuition;
  
• understanding difficult ideas;
  
• explaining mathematics to others;
  
• deciding which problems and theories are interesting;
  
• building mathematical communities.
  

In his view, these activities may become more valuable precisely because automated systems can increasingly handle the mechanical production of mathematical text.

3. The PhD should change dramatically
One of the most concrete proposals concerns mathematical PhDs. Litt argues that a thesis can no longer reliably demonstrate that its author understands the mathematics it contains, because AI can increasingly generate substantial mathematical text. Instead, he proposes reconceiving the PhD as demonstrating that the student has become a world expert on an interesting, deep topic and can explain and defend that understanding. 
A thesis could still exist, but the decisive assessment would be a rigorous oral defense. Students might use AI extensively in their research—or not use it at all—but they would have to demonstrate genuine understanding.
He also suggests more frequent situations where students must independently solve unfamiliar examples or apply techniques in new contexts. In other words, the ability to understand mathematics should become a stronger signal than the ability to produce mathematical prose.

4. Papers may become less important as evidence of expertise
This leads to a broader institutional argument. Hiring, graduate admissions, promotion, and other academic decisions currently rely heavily on papers and other written outputs. Litt believes this system will become increasingly unreliable as AI-generated mathematics becomes abundant. Talks, interviews, seminars, and sustained mathematical discussion may become better ways of determining whether someone actually understands their work.
Importantly, he does not argue that AI-generated mathematics should be rejected. Instead, he suggests separating two questions:

Is the mathematics good?
Does the human presenting it understand it?

AI could legitimately contribute to the first without automatically satisfying the second.

5. The community's ability to decide what is interesting becomes crucial
A particularly interesting part of the essay concerns mathematical taste.
AI might eventually generate enormous numbers of conjectures, proofs, theories, and research programs. But mathematics cannot simply value everything equally. Someone—or some community—must decide which questions are profound, which ideas are fruitful, and which results deserve human attention.
Litt suggests that mathematics should therefore explicitly reward people who identify important research directions and persuade others that those directions matter.
He acknowledges that AI itself may eventually become very good at asking interesting questions. Nevertheless, he believes AI will not necessarily ask the same questions humans would ask, leaving room for human mathematical culture and judgment.

6. Mathematics could become dramatically cheaper
Another major change is economic. Historically, many mathematical questions required enormous amounts of human time. AI could convert some of that cost into a relatively small monetary expense.
Litt imagines situations where a difficult mathematical question might be answered for roughly the cost of an expensive dinner. Rather than seeing this as the destruction of mathematics, he argues that we should celebrate the ability to answer basic questions cheaply—because every answer creates new questions. 
A simple AI-generated solution might therefore lead to:
answer → deeper question → confusion → seminar → research community → new mathematics
The important human contribution becomes the effort to understand what the machine has produced.

7. We cannot simply stop AI mathematics
Litt is skeptical of strategies based on hoping that AI labs will abandon mathematics or that technological progress will somehow reverse. He argues that even if particular AI companies disappeared, the underlying technology would remain. Consequently, mathematics needs an institutional strategy that works even if increasingly powerful mathematical AI is inevitable. 
He also rejects the idea that mathematical problems can simply be “reserved” for human researchers. If people can access powerful AI systems, they will use them. Instead of attempting to prevent mathematical experimentation, Litt advocates creating a culture capable of absorbing the resulting abundance.

8. The article's most important philosophical point
The essay ultimately distinguishes between mathematical production and mathematical understanding.
AI could potentially produce mathematics at a scale and speed vastly beyond humans. But, according to Litt, no machine can understand mathematics for us. A student can ask an AI for an explanation, receive a beautiful explanation, and still not understand it. The human being must perform the cognitive work of understanding. 
This gives the article its optimistic conclusion: mathematics is not running out of things to discover. Even if AI solves enormous numbers of problems, the space of things humans can understand remains effectively infinite.

“We’ve always been at the beginning, and we always will be.” 

Overview
The most important idea in the article is not that AI will replace mathematicians, nor that mathematicians will simply continue doing mathematics as before. Litt is proposing a change in what counts as mathematical expertise.
The future mathematician may be less valuable for producing a proof from scratch and more valuable for being able to ask profound questions, recognize important mathematics, understand AI-generated results deeply, connect apparently unrelated theories, explain difficult ideas, and build communities around them.
That is a significant shift from “Can you prove this?” toward “Do you understand why this matters?”
And that distinction may become one of the central questions for mathematics education, PhD training, research institutions, and mathematical culture over the next decade. 

ARTICLE