When AI Research Assistants Turn Into Competitors
Mathematicians shared unpublished drafts with Codex. Shortly after, OpenAI announced a breakthrough on fluid equations. Research assistants may now become research competitors.
No evidence currently proves the private drafts were used to produce this solution. But “we did not search user data” does not answer the separate question: was the material used for model training?
Two researchers spent roughly a year collaborating privately on the problem. Meanwhile, OpenAI reported deploying around ten thousand AI agents; they obtained results in 88 hours, then spent another 17 hours on formalization and verification.
Timelines overlap and research directions align, enough to spark questions — yet not enough to prove one side appropriated work from the other. The 88-hour window starts when the agent task launched, not a countdown triggered by draft uploads. The critical thing to verify is what happens to a draft once it enters the assistant system.
The pair shared unfinished work with the AI before formal publication.
NYU mathematician Tristan Buckmaster and his collaborator Levent Alpöge spent about a year studying fluid equations. Alpöge works at another AI firm, and this was their personal joint project. Over the course of their research, they fed successive unpublished drafts into OpenAI’s Codex‑S. For them, AI was already embedded in the exploratory phase, long before a final paper existed.
On September 8, 2026, OpenAI released its result on the Navier‑Stokes equations. These are the equations governing fluid motion. The firm claimed to construct finite-time blowup under smooth external forcing. Blowup here means the mathematical solution develops a singularity within finite time, while the forcing term itself remains smooth; the result cannot rely on artificially inserting singular forces. Whether this proof stands up still awaits scrutiny from the mathematics community.
OpenAI’s timeline: on September 1, researchers heard rumors of a related breakthrough and launched roughly ten thousand parallel AI agents to tackle the problem. Results emerged after 88 hours, followed by a further 17 hours of formal proof and validation. Important distinction: the 88-hour clock starts when agents began running, not immediately after draft submission.
Parallel discoveries naturally raise questions about provenance, but matching timing and topic alone cannot establish plagiarism or theft. Buckmaster himself publicly stated he did not know whether his private data had been used.
OpenAI’s response: Chief Research Officer Mark denied that anyone specifically searched user data for this solving effort. The official post also stated OpenAI researchers and AI agents had not seen the pair’s work via any channel before their results were made public.
Three distinct scenarios must be separated:
- You upload a draft to a cloud AI assistant; the cloud processes the content to answer your query — this is real-time tool response.
- A human investigator manually retrieves your old draft to use in an unrelated internal project — this is a separate access event.
- Content from your draft is selected into a training corpus to refine the base large model — this is a third, entirely independent pathway.
Denying “targeted searches of user data” logically does not answer whether the material entered model training. Even if the text was used to refine the model, that still does not directly prove it guided the AI to solve this problem; solid evidence would be required for every link in that chain.
Could reviewing the prompts used during solving clarify matters? OpenAI said it offered the two scholars full access to the task prompts and the completed proof itself. This lets reviewers check what instructions were given to the AI. However, examining inference prompts reveals nothing about what material the model ingested during training. Model inputs at runtime and training data seen long before are two separate things to verify.
Without access to full training datasets and system access logs, researchers cannot independently trace every system path their unpublished drafts travelled through.
There remains a plausible alternative: OpenAI simply leveraged stronger compute and reasoning capability and arrived at the result entirely independently. Public materials do not rule this out.
Even if independent breakthrough is confirmed, the underlying boundary issue remains urgent for researchers.
The new reality: the platform acting as your research assistant may also run competing research in the same field. Tool providers become potential rivals.
In this setting, an unpublished draft carries value beyond a complete final answer. Even a single promising core idea inside the manuscript can save the party that sees it years of dead ends, which can then be scaled up with massive compute. This is a risk the research community must guard against.
There is still no evidence this risk materialized in this specific case. It is also important to avoid oversimplifying rigorous mathematical proof. Even with access to a promising line of reasoning, the formal proof work remains extremely challenging; solving hard math cannot be reduced to letting machines run code.
Data policies differ across products. OpenAI Enterprise and API calls default to excluding user data from training. Personal accounts also have a toggle to opt out of training usage. Any discussion of these drafts must reference the exact product and settings used by the researchers, not blanket assumptions that all cloud uploads automatically enter training datasets. Policy text defines permitted scope, but actual events depend on retained system logs.
About the 17-hour verification phase: OpenAI used computer formal proof tools. These tools only check whether the full derivation remains logically consistent under given definitions. They perform no attribution traceback to identify which human manuscripts shaped the model before the run.
So while AI’s conquest of famous math problems deserves awe, questions over data provenance remain fair to ask.
For the mathematician who uploaded drafts: AI research assistants are powerful. But trusting the platform with raw, half-formed ideas requires knowing exactly where those ideas go and how they may be reused. And when disputes arise, researchers need mechanisms to audit what happened.
If you enjoy what I do, consider supporting me! Every little bit means the world! ❤️
☕ Buy me a coffee