The fear of “AI obsoleting mathematicians” has become another obsession by the media over the past 2 years, along with everything else AI related, with new headlines appearing every week:
For example, this article recently went viral, “What it Means to Be a Mathematician When AI Does the Math?” Like the rest of academia, mathematics is not immune to gatekeeping, fads, cliques, and status-signaling. If the profession of mathematics were to disappear because of AI, it would not represent a loss to humanity. The act of doing math would continue–especially in the private sector such as developing algorithms–but it would no longer exist under the formal title or appellation of “mathematician.” It would just be “people doing math”.
People always get mad when you discuss this, on either side of the debate. It’s much more personal than debates about sports teams. Supporting the Patriots is not an indictment on one’s moral worth or intellect, unlike about AI. If you’re pro-AI you’re accused of devaluing the artistic or aesthetic value of art or other expressions of creativity. If you’re skeptical, you’re accused of being “a doomer,” a luddite, or “just not being smart enough to understand it”.
But the good news for mathematicians, is I predict they will not be obsoleted by AI. What is lost in the hype is no AI can tell you if a proof is correct. An AI can produce a convincing looking proof, but it can have a subtle but critical error or make an assumption that is unfounded.
You can test this yourself with Claude and Chat GPT. I had Claude write a proof. I fed it into GPT, which flagged errors and logical gaps with the proof. I asked GPT to write a summary, which I fed into Claude. Claude made adjustments, which I again fed into GPT. GPT flagged additional mistakes. I repeated the process until both ran out of tokens. After at least 6 attempts at refinements, neither could actually definitely prove it to the satisfaction of the other. This is true for all but the simplest of problems or concepts.
Thus, it ultimately comes down to humans as the sole arbiter of correctness. A mathematician has to craft the prompt, and another mathematician to interpret/check the results. Even with LEAN, a mathematician still has to write the code and evaluate the output. LEAN proofs are often extraordinarily complicated and verbose, requiring thousands of lines of code for difficult results, and the system only works for a limited subset of mathematical problems.
Also, these programs are very expensive and propitiatory. They are not like the commercial AI that regular people use. It takes considerable prompting and trial an error to solve even Olympiad/Putnam problems, and tons of work by humans pouring over the results to see if it’s correct. For every Erdos problem that captures the headlines, there are many where it failed or untold hours of prompting and token burn to get that result, and manhours verify it.