doomslide

I think AI will expand mathematical discovery, especially where proofs are too combinatorially complex for unaided human search. But a modest productivity boost is not the same as moving the frontier: the transformative case depends on systems finding proofs or structures we would not realistically find ourselves. The institutional consequences may matter more than raw intelligence. If the models, compute, data, and evaluation scaffolds remain controlled by a few companies, then capability becomes difficult to verify and mathematical knowledge risks moving from public papers and discussions into proprietary chat silos. That is a route to disempowerment through concentrated infrastructure and ownership, not necessarily through machines becoming intellectually supreme. It also creates an attribution problem: companies can claim discoveries while withholding enough of the process that outsiders cannot properly audit what happened. Open access could change this trajectory and make mathematicians substantially more willing to adopt these tools. Formal languages such as Lean may also offer a way to make outputs checkable and perhaps constrain model behavior programmatically, but that is a conjecture, not an established alignment solution. So my expectation is neither “AI kills mathematics” nor “AI simply accelerates it.” It changes who can discover, who can verify, and—most importantly—who owns the resulting knowledge.

Overall, I expect a mixed technological gain coupled to a negative institutional default. AI will likely produce genuinely valuable mathematical discoveries, particularly through large-scale search, but those benefits do not automatically translate into broadly shared progress. If the relevant models, compute, data, and evaluation machinery stay concentrated, the likely result is greater dependence on a few firms, weaker public verification, and mathematical knowledge leaking from shared discourse into private interfaces. Intellectual-property arrangements worsen this by letting companies own the compressed machinery built from culture while creators and users bear the costs. So the decisive variable is not capability alone but access and control. Open models and public, formally checkable outputs could make the impact substantially better. Without that, I expect real discoveries inside an increasingly oligarchic knowledge system.