Daniel Lemire, a computer science professor at TELUQ Université du Québec, posted a provocative take this week on the anxiety spreading through mathematics departments as agentic AI improves. His argument is blunt: if a top high school student with a large token budget can produce the equivalent of a good PhD thesis as a hobby, the problem is not the AI. The problem is that mathematics, like much of academia, has been using research output as a sorting mechanism for jobs rather than as a means to solve problems.
The panic and its source
Mathematicians are reacting to the prospect that AI systems can now generate proofs and research at a level that previously required years of graduate training. Lemire frames the concern directly: if someone completed a PhD thesis in mathematics in 2022, a sufficiently capable AI can now produce equivalent output while a smart high schooler plays video games. The implication is that the years of effort, the qualifying exams, the advisor relationships, and the publication record that once served as signals of ability are becoming reproducible by machine.
But Lemire does not frame this as a loss. He frames it as a correction.
The sorting mechanism problem
The core of the argument is that academic research lost its external audience decades ago. Peer review, introduced widely in the 1970s, turned research output into a closed loop: researchers write papers, other researchers review them, journals publish them, and committees use publication records to decide who gets jobs. The papers themselves are rarely read outside the subfield. The purpose of the output is to sort people into a hierarchy, not to solve problems that matter to anyone outside the academy.
The numbers support this reading. Holding a PhD today gives roughly a 10% chance of securing a tenure-track or tenured position, and that number is falling. The supply of PhDs has been deliberately increased for decades, while the number of positions at the top has not grown proportionally. The result is a system where the majority of PhD holders are competing for a shrinking fraction of the jobs their credentials were supposed to lead to.
AI disrupts this sorting mechanism because it makes the output, the paper, reproducible. If the paper was the signal of ability, and AI can produce papers, the signal collapses. Lemire's point is that this is not a new problem caused by AI. It is a preexisting problem that AI has made impossible to ignore.
What the focus should shift to
Lemire published a piece in Communications of the ACM in September 2024 arguing that AI's ability to generate research text should push the scientific community toward evaluating impact and problem-solving over publication volume. His position this week extends that argument to mathematics specifically. Nobody wants a paper about cancer, he writes. We want to eradicate cancer. The emphasis should be on results, not on the process of passing peer review, which AI can now replicate.
The prediction is that over time, the final output of research will stop being the paper. It will become the solution, the tool, the dataset, or the changed outcome. Papers will be intermediate artifacts, not endpoints. This is already true in parts of computer science and engineering, where working code and deployed systems carry more weight than publication records. Mathematics has been slower to move in this direction because proof is the natural output, but the argument is that proof for its own sake, disconnected from application, is exactly the kind of output that AI can now commoditize.
The math-can't-cure-cancer objection
Lemire anticipates the pushback: mathematicians work on abstract problems, not on curing cancer or building antigravity machines. His response is a challenge: maybe it is time they try. The point is not that pure mathematics is worthless. The point is that a field whose primary output is papers, and whose primary audience is other mathematicians, is the most vulnerable to disruption by a system that generates papers. Fields with external outputs, buildings that stand, drugs that work, software that runs, have built-in resistance because the output is tested against reality, not against peer reviewers.
Intelligence is not a scalar
The final thread in the argument addresses the fear that AI will replace researchers entirely. Lemire says it has not happened for him despite extensive use of AI tools. His framing is that intelligence is not a single dimension. Once an AI can do one thing, humans find more work to do. The displacement is real, but it is displacement of specific tasks, not of the need for people who understand problems and can direct tools toward solutions.
The implication for developers and researchers is practical. If your value comes from being able to do something that AI can now do, the margin is shrinking. If your value comes from understanding which problems are worth solving and directing tools toward them, the margin is expanding. The distinction matters for career planning, for hiring, and for how institutions evaluate output. The fields that adapt first will be the ones that define what counts as a result, rather than waiting for AI to make the current definition obsolete.