Here’s a fun game: take months of grueling, caffeine-fueled intellectual labor—say, a breakthrough in topology or a brand-new proof—and paste it straight into a proprietary chatbot owned by a company desperately hunting for high-value training data.
Sounds like career suicide, right? Yet, that’s essentially what modern researchers keep flirting with, and I’m frankly baffled we even have to talk about it.
Over on the fediverse, mathematicians like Andreas Thom have been sounding the alarm again, asking the uncomfortable questions that the Silicon Valley hype machine would rather bury. Can we actually trust OpenAI with unpublished math? Spoiler alert: no. You shouldn’t trust a black-box LLM with your grocery list, let alone the culmination of your academic career.
Let’s cut through the techno-utopian fog for a second. OpenAI isn’t a friendly research collective run by kindly uncles who love numbers. It’s a commercial behemoth that survives on one thing: consuming human intellect to make its next model marginally better at hallucinating poetry. When you feed an unpublished theorem into their ecosystem to check your work, look for errors, or rewrite it in rust for some absurd systems-level simulation, you aren’t just using a tool. You’re handing over the keys to your intellectual kingdom.
Sure, the terms of service say they won’t look at your enterprise data. Right. And big tech has never quietly shifted the goalposts when a new model needs training fuel. History is littered with the corpses of creators whose work was vacuumed up because of a vague clause in an updated privacy policy. If it’s digital, and it passes through their servers, treat it as public domain. It really is that simple.
What gets me is the sheer naivety in academia. We’re talking about some of the smartest people on the planet—minds that can untangle abstract dimensions that make ordinary brains melt—yet they fall for the same productivity trap as mid-level marketing managers. “It’s so fast!” they cry, as the bot spits out a plausible-looking proof that quietly inverts a foundational axiom. And worse, even if the math is right, you’ve just trained the corporate oracle on your secret sauce.
If you want to use AI for math, fine. Run an open-source model locally on your own hardware where you actually control the bits and bytes. Stop outsourcing your paranoia to companies whose entire business model relies on taking your stuff without asking.
Here’s my hot take: If you hand your unpublished theorems to OpenAI, you deserve to get scooped by an algorithm trained on your own brilliance. Stop feeding the beast that’s eventually coming for your job.