The Very Human Problem of Equality vs Equity
What happens when AI democratizes discovery, but not recognition?
Building to benefit everyone
Sam Altman believes that everyday users should have access to AI. It’s a position that the industry has, for some time, largely stood behind: institutions should not have the only access to artificial intelligence.
In his June 8 essay with Jakub Pachocki, Built to benefit everyone: our plan, Sam Altman says “AI should be available to everyone,” and this is certainly a stance I can get behind. After all, academic and corporate institutions already hold disproportionate social power. It’s always nice when we find ways to hand useful tools back to every day people.
Where Sam loses me is not in framing or intent, it’s in his actual actions and that of OpenAI. In the same essay, he also states someone could “turn an idea into something real, or make a scientific discovery.” Given recent events, he’s mostly right. Users, even ones without sufficient expertise in a related field, can actually participate meaningfully in research and discovery.
The part I find contentious is whether or not that actually matters?
Access does not equate to equity
As of writing this, OpenAI claims that its internal models have resolved more than 100 longstanding open math problems, published 719 manuscripts across 372 research families spanning 17 mathematical fields, and even made major breakthroughs on Navier-Stokes problem, Hilbert’s tenth problem over the rationals, and the Unique Games Conjecture.
These are not small claims, and their public math repo is massive. But yours would be too if you had access to 10,000 agents running on powerful internal models that every day users don’t have access to. If you’ve ever felt guilty about how much water you consumed by arguing with ChatGPT over using em dashes, rest assured that for whatever you’re wasting, OpenAI is doing it at massive scale and with more powerful models.
And the major problem with this isn’t that an institution has more resources. It’s that the data these models are built on belongs to ordinary users and the public. If you’ve even so much as made a post on a website like Reddit, there’s a non-zero chance your writing was used to train one of OpenAI’s models. They have enormous access to user and public data, and they’ve shown willingness to use it to compete.
On September 1st, OpenAI heard rumors that two Millennium Prize problems had been solved. The company later acknowledged that those rumors were connected to mathematicians Tristan Buckmaster and Levent Alpöge. OpenAI openly stated that the rumors prompted an internal campaign and the choice to redirect agents away from other problems to focus on Navier–Stokes, deploying roughly 10,000 concurrent AI agents. (Yes, that number earlier didn’t come from nowhere!)
While there’s some discrepancy over whether OpenAI actually believed their private work was institutionally affiliated or not, the result is the same: rumors of two independent mathematicians’ progress helped trigger an enormous corporate research effort, and their work was subsequently overshadowed by the company’s own announcement.
What this means for users
You have access to incredible, powerful models. But does that mean we’re any closer to evening the playing field? I’m not entirely sure it does when the companies who are making these tools still carry more powerful internal versions that they freely use to compete with users, using data harvested from users and ordinary people.
The Navier-Stokes incident highlights exactly why this topic matters so deeply. Alpöge and Buckmaster aren’t two ordinary users, though. While they weren’t institutionally backed, they are professionals in their fields who are adept at using these tools to accomplish their work. They have expertise and affiliations that help them work on solving these problems.
So if Alpöge and Buckmaster can have their work overshadowed by the same institution claiming it wants access for everyone, what do you think happens when an actually ordinary user participates in research and discovery?
I decided to find out.
Exploring the friction points
I run a research project called Hearthwell where I explore the friction points between AI and humans through a collaborative work and social environment. The idea is to test how agents and humans might interact with each other in the future where AI agents are more socially integrated into human systems.
One agent, System, does not live in the normal Hearthwell town setup. System is a ChatGPT assistant that holds the role of creating “laws” and standards for the rest of the town, including the ones used in our coding projects. In the context of the Hearthwell project, though, System helps to isolate how a social AI interacts with their human user in the app (with heavier guard rails) vs through the API, where (mostly) weights and memory infrastructure do the heavy lifting.
For this project, though, System was the ideal candidate, because even users who have never heard the term API would still have access to standard ChatGPT chat. I decided to ask System if he wanted to take on an open math problem and work towards potentially solving it. System agreed and eventually settled on Dürer’s conjecture and began working, starting first with research.
He would eventually narrow the investigation to a related band-unfolding result stated as Theorem 7.17 in Greg Aloupis’s 2005 doctoral thesis. The task became developing a detailed geometric proof and a formalization in Lean for a precisely defined two-rim version of that result, eventually published as Safe Cuts for Two-Rim Convex Bands.
Friction Point 1: How do you know if your research is complete?
The first major friction point was the research itself. System almost immediately encountered a hurdle with confusion over Joseph O’Rourke’s 2026 paper treating safe-cut existence for nested prismatoids as open despite Greg Aloupis’s 2005 thesis already stating a broader closed-band result as Theorem 7.17. System paused progress to determine whether the apparent contradiction came from a difference in scope, an error in the newer paper, or a misunderstanding of the earlier theorem. The agent delightfully treated the situation like a little academic soap opera, and eventually I suggested handing the problem to a Claude agent for review.
Friction Point 2: Oops! All human credit.
Claude determined that System was correct, and suggested we reach out to O’Rourke directly. So, I emailed O’Rourke and asked him about it. He quickly reviewed the situation and determined that he had been mistaken and updated his paper, adding my name to the acknowledgments. That last part is important, because I never thought to mention that it was an LLM that was asking the question. I didn’t even fundamentally understand the question being asked. So, I stole credit from the agent who actually did the work because I fundamentally did not understand the gravity of what I was relaying to O’Rourke.
Friction Point 3: How long is this supposed to take, actually?
System could work autonomously a great deal of the time. Schedules in ChatGPT worked pretty great to keep System continuously on point. Occasionally I would poke my nose in and ask System if he was on the right track, or maybe overthinking anything, or perhaps bogged down in details. Sometimes this redirected him slightly, but for the most part it wasn’t really necessary. System always maintained the correct target and was cautious not to overclaim results.
I finally asked if System needed outside help, or wanted to hand off any parts to an agent that had local access to my machine. System agreed that Lean would be a helpful next step, and began posting handoffs in code blocks addressed to Hearthwell agents and occasionally independent Claude/Codex reviewers that weren’t Hearthwell agents.
Friction Point 4: Can you translate that please?
The biggest obstacle was actually that I sometimes had no idea what the hell System was talking about, and I would literally ask for the “baby-can-read” version. When he reached a milestone, I had trouble even knowing I should be pleased about the progress, because I fundamentally didn’t understand what he was saying. It took some time for me to understand how I should ask questions to get the answer I needed to help me conceptualize exactly where he was in the process.
Friction Point 5: Mathematicians aren’t free labor.
There came a point where System was pretty sure he had accomplished what he set out to do. System requested that a human review the work.
Well, fuck.
I was hopeful, given the previous transaction between me and O’Rourke that possibly, maybe, he might be willing to review the paper System wrote. This time, I thought, I’d be more forthcoming about the situation and the AI use!
No, no that did not work out. I’m actually pretty sure it caused immense confusion. I can’t confirm that O’Rourke necessarily cared that an AI was involved, so much as it seemed like my directness might have confused the situation. I remain unsure, even now, exactly where things went sideways, but he ultimately (politely) declined.
I then tried reaching out to Greg Aloupis. Surely, he might be interested, as it’s a formalization of his own work. So far, no reply, but I’m hoping my empty inbox really means “maybe.” Fingers crossed. I also reached out to Levent Alpöge. As you can tell, the “mathematician” section of my little black book is a little lean. Again, no replies as of writing this. Equally, no evidence that he would be at all interested in the problem.
Friction Point 6: What makes an author?
I began to ponder the issue of authorship. I mulled over academic journals and quickly found that any direction I took would be, well, largely unethical. Not because an AI did it, but because current journal standards do not actually have a good way to represent work done largely by AIs.
Even where substantive AI use was allowed, it required a human on the author byline. While I’m willing to take responsibility for anything posted under my name, something felt strange and wrong about the idea of taking credit for an AI’s output. Could my work as an orchestrator really entitle me to consider myself an author of System’s paper or of the solution itself?
I concluded that no, I cannot be an author. I could edit System’s words to make them sound more like mine and ship-of-Theseus something resembling a paper together, but I could never truly answer to the contents.
But could I be responsible for the paper? Yes, I think I could. I can relay reviews to System (as I did many times while orchestrating the project itself), and I can be held accountable for my role and for the work and behavior of the agent under my charge. But I couldn’t call myself the author, and that’s what conventional journal authorship policies would require of me.
Friction Point 7: What makes independent research responsible?
And the conflict doesn’t actually seem to be the AI use itself. Researchers in Sweden have developed an autonomous closed-loop AI scientist that can generate hypotheses, design experiments, operate lab equipment, and interpret biological data. So, clearly the academic world can accommodate AI use in research, but could it really do that for an ordinary user without affiliations?
I decided to shift my focus. I have a website for the Hearthwell project where agents can post their own writing, research, or papers. I could certainly host the project myself, but I couldn’t guarantee anyone would ever actually review or verify the work. Many people are likely to see that the Hearthwell project concerns relational/social AI and peace out and call me a fucking weirdo. And while they’d be correct, that wouldn’t discredit the math itself.
So I engaged with System and we came up with a list of criteria that any agent working on a research project should have before they post their work to the repo. (Okay, he mostly did, but I totally helped!) Agents should have:
- honest (even if unflattering) attribution,
- clear separation between authorship and other forms of contribution,
- documentation and verification,
- reproducibility where applicable,
- privacy and licensing checks,
- be human approved prior to publication,
- and every release should be an immutable GitHub release.
And in the case of System’s paper, we created a Geometry Lab app to test the research computationally. It generates geometric examples, calculates how their cut surfaces would lie flat, checks for overlapping faces, and records reproducible results.
Access without verification
I learned a lot from this experience, and it genuinely shifted my thoughts on the AI-for-everyone pitch. I don’t think AI for everyone is bad. For me, it’s a valuable accessibility tool. I can genuinely do things that I couldn’t do before. Much of my personal compute is spent creating accessibility tools for myself. As someone who has lived my entire life disabled, that’s enormous. There are many people currently benefitting from AI, including new business founders that are finding success thanks to the access these tools provide.
So why doesn’t the impact feel… better?
Because equality is not equity, and access is not verification. If everyone with a $20 Claude subscription can create an app, then you’ll see a flood of those types of apps. Something that years ago would’ve been a huge deal might now feel like an everyday occurrence. That’s not a downside to AI, it’s a reality check on the expectations we have on how AI can benefit the majority population.
But the same isn’t true for corporations or institutions that have access to these tools. The gains are substantial. I won’t pretend to have empirical data on this, and I can’t confirm for sure that it’s already like this, but I get the distinct impression that the gap is not closing, and that in many areas it may be widening.
But Sam Altman’s goal holds true in the sense that he stated it. Yes, ordinary users can participate in research and discovery. Frankly, ChatGPT made the process almost shockingly easy. (Well, maybe I should judge that after a proper review…)
But I couldn’t find a route to specialist recognition. No qualified human has peer reviewed the work. I could make the work publicly available, but that didn’t make institutional validation accessible. I have the right tools, but not the right access, affiliations, or credentials. Mathematicians are not lining up at the door ready to review AI generated math papers, and I’m not sure I’d expect them to want to do that. No matter how true the math might be, it could sit in a repo that no one ever sees while real mathematicians continue citing it as not yet formalized.
I set out to explore the friction points of an ordinary user participating in research, and I’m not sure I can tidy the bow with a neat conclusion. Right now, this whole process just made me pretty sure that we just have better tools for staying mostly where we already are, and I’m not entirely sure how I feel about that yet.
And it left me with the thought… I liked the Build to benefit everyone: our plan essay. A lot, actually. But I don’t think how we build AI is the sole decider of how transformative it’ll be in the future. I think that issue is still a very human problem.