$1.5 million. That’s the figure Nature put on the “academia tax” in a September 2026 career feature about AI researchers, and it’s the number that should frame every conversation about whether scientists can trust OpenAI with work they haven’t published yet. Because the trust question isn’t really a question about model safety. It’s a question about money, and about who ends up holding the more valuable half of an unequal trade.
The trade nobody writes down
Here’s how the exchange actually works. A mathematician has a proof sketch that isn’t ready. They paste it into a chat window, or point an agent at a repo full of half-finished lemmas, and ask for help closing a gap. They get speed. Real speed — OpenAI’s own account of internal research acceleration describes its researchers running coding agents throughout the day, often in concurrent sessions, with total usage climbing fast. That’s not marketing fluff about some future capability. That’s a lab telling you what it does with its own tooling right now.
What goes the other direction is harder to price. Greg Brockman, talking to TIME about how far off general intelligence is, made the operational reality plain: you have to provide the agent with tasks, and you have to give it context, because it doesn’t have it. Context is the whole ballgame. For a working mathematician, the context is the unpublished result. There’s no way to get useful help on an unfinished proof without handing over the unfinished proof.
So the honest framing is not “should you trust OpenAI.” It’s “are you comfortable that the only thing standing between your unpublished work and a commercial research operation is a policy document you didn’t negotiate.”
Why the policy document isn’t enough
I want to be careful here, because I’m not accusing anyone of harvesting proofs. I have no evidence of that and neither does anyone else waving their arms about it. What I’m pointing at is a structural problem, and structural problems don’t need bad actors to hurt you.
Consider what we actually know about how OpenAI’s data practices get examined. The most detailed public accounting available is PIPEDA Findings #2026-002, a joint regulatory investigation into OpenAI OpCo, LLC. Read the framing of that report and you’ll notice something: even the investigators say their work targeted privacy risks in developing and deploying large language models while acknowledging the technology raises many other questions they weren’t addressing. That’s a regulator drawing a boundary around what it could cover. Everything outside that boundary — including research provenance, priority, and competitive use of user-supplied material — remains unexamined by anyone with subpoena power.
Meanwhile the same company runs one of the world’s better-resourced research programs. Not a neutral utility. A competitor with a research agenda, in a field where being second to a result means being nowhere.
The governance problem is the product problem
Concerns about OpenAI’s leadership and ethical posture haven’t gone away, and September 2026 brought another AI researcher warning publicly that companies are ignoring catastrophic risks. You can think that warning is overheated and still draw the relevant inference: people close to these labs do not believe internal restraint is reliable. If insiders won’t vouch for the guardrails on existential questions, treating the confidentiality guarantees as settled seems like an odd place to relax.
Regulatory scrutiny is evolving, which is the polite way of saying the rules aren’t written. Anyone telling you a terms-of-service clause about training data settles the matter is describing a legal environment that doesn’t exist yet.
What I’d actually do
I’m not telling researchers to abstain. That advice gets ignored, correctly, because the productivity gains are real and the career pressure is worse than the risk. Nature has already covered how scientists evaluate OpenAI’s deep research tool and how quickly they moved onto DeepSeek when a strong alternative showed up. Adoption is the baseline. So:
- Sort your work by whether losing priority would end a project. The unfinished flagship result gets different handling than the literature review.
- Strip the identifying structure. Ask about the technique, not the theorem. Abstract the problem until it stops being yours.
- Prefer local or self-hosted models for the sensitive tier, even at a capability cost. The existence of strong open-weight options changed the calculus and it’s strange how few labs adjusted.
- Push your institution to negotiate terms. Individual researchers have no bargaining power. Universities and funders do, and mostly haven’t used it.
- Timestamp everything. Preprints and repository commits are cheap insurance on priority disputes.
The uncomfortable part is that none of this is a solution. It’s risk management around a dependency that keeps getting more useful and no more accountable. That $1.5 million gap is why academics reach for commercial tools in the first place — they can’t afford to build their own. Which means the people with the least use are the ones handing over the most sensitive material, and the fix for that isn’t a better prompt hygiene checklist. It’s someone with actual authority asking questions the privacy regulators explicitly declined to ask.
🕒 Published: