Unpublished math is a fragile asset.
It exists as a claim nobody else has made yet, and its entire value sits in that gap. So when a researcher pastes a half-finished proof into a chat window to ask for a sanity check, they are doing something categorically different from asking a model to rename variables in a Python file. That distinction is getting blurry fast, and the questions about whether OpenAI is a safe place to park that kind of work keep coming back.
I build bots for a living. My instinct here is not to argue about corporate character. It is to look at the plumbing.
The usage curve is the real story
OpenAI has described how its own researchers’ daily work shifted over the past year: coding agents running throughout the day, often in concurrent sessions, with total usage climbing quickly. Read that as a signal about the shape of adoption everywhere, not just inside one lab. Scientists have flocked to these tools for the same reason developers did. Nature has covered both the deep research tool’s usefulness for scientists and the rush toward DeepSeek. The demand is not speculative.
Concurrent agent sessions change the trust calculation in a way single-turn chat never did. One prompt is a decision. Twenty agents running in parallel across your repo, your notes, and your draft manuscripts is a policy, and most people never wrote that policy down.
Context is the attack surface
Greg Brockman, talking to TIME about the road to general intelligence, described the current reality of agent work as micromanagement — you hand the agent tasks, you supply the context, because it does not have that context on its own. He meant it as a limitation to be solved. For anyone guarding unpublished results, it is also a precise description of the risk.
Agents are only as useful as the context you feed them. Useful context for a mathematician means the actual lemma, the actual failed approach, the actual thing not yet on arXiv. There is no way to get high-quality help on original work while withholding the original work. That tension is structural, not a policy toggle.
Regulators are asking the same questions differently
The PIPEDA joint investigation findings into OpenAI OpCo, LLC landed as a privacy document, aimed at the risks tied to developing and deploying large language models. Notably, the report itself acknowledged that the technology raises many questions beyond the privacy scope it addressed. That is a regulator saying, in careful language, that the frame is incomplete.
Scrutiny keeps evolving, and concerns about leadership and ethical risk have not gone away. One AI researcher recently warned publicly that companies are ignoring catastrophic risks. None of that tells you whether your proof draft is safe. It tells you the accountability structures are still being built while adoption races ahead.
What I’d actually do about it
Trust is not binary, and treating it that way leads to bad engineering. Here is the practical version for anyone wiring research work into these systems:
- Tier your inputs. Published results, tooling code, and literature review are low-stakes. Unproven claims and novel constructions are not. Route them differently, with different tools and different retention settings.
- Log what leaves. If you cannot answer “what did my agents send out last Tuesday,” you have no basis for trusting or distrusting anything. Build a proxy layer that records outbound context before you scale up sessions.
- Keep an open-weight fallback. The scramble toward DeepSeek showed how quickly researchers will adopt an alternative. A locally hosted model you can run on your own hardware handles the sensitive tier at lower capability but with a clear boundary.
- Separate the agent’s workspace. Give coding agents a scoped directory, not your whole filesystem. Concurrent sessions with broad access are how sensitive files end up in prompts nobody intended.
- Write the policy before you need it. Lab-level rules about what goes into a hosted model beat individual judgment under deadline pressure.
The cost question underneath it all
There is a reason this debate has teeth. Nature’s reporting on the $1.5 million “academia tax” points at the squeeze: compute is expensive, hosted tools are the affordable path, and researchers who opt out on principle may simply move slower than the ones who do not. Trust discussions get shaped by budgets, and that is uncomfortable but true.
My read is that the answer is not abstention or blind adoption. It is instrumentation. The people building bots learned this the hard way with API keys and customer data — you do not trust a pipeline you cannot inspect. Unpublished math deserves the same discipline as production credentials, and right now it mostly is not getting it.
đź•’ Published: