Is OpenAI Taking Everyone for Fools?

On the same day two academic researchers publish their findings to a mathematical problem, OpenAI claims to have used their AI model to come to a very similar conclusion. Not only this, OpenAI acknowledges it started the effort to solve that problem only after it became aware of the researchers’ work. That is a lot of coincidence.

The story, briefly

So, let’s briefly look at what happened:

On 8th September 2026, Tristan Buckmaster, Professor of Mathematics at the New York University, published a statement saying that “today, Levent Alpöge and I have made public three results: finite-time blowup with smooth forcing for incompressible porous media, for Boussinesq, and for 3d incompressible Euler” (Buckmaster 2026).

In layperson’s terms, they present a solution to a problem that had kept mathematicians busy for decades. The mathematical details of this achievement are not relevant here. Unfortunately, there was little mood for celebration: “There is another part of this story”, Buckmaster writes, “and one that, honestly, I very much wish I did not have to be concerned with”.

He explains that the week before “a rumor [was] circulating […] with Levent having received tips that information about our progress had been passed to OpenAI” and that “an internal OpenAI model had produced a proof […] for the forced Navier-Stokes equations.” As Michael Harris, Columbia University professor, points out, this strongly suggested that OpenAI “got wind of what Buckmaster and Alpöge were doing and decided to scoop them” (Harris 2026).

When Buckmaster contacted OpenAI about this, OpenAI “twice asserted that [they] wanted Levent removed from authorship”, indicatingsuggesting the reason was that Alpöge worked with Anthropic. Graciously, OpenAI offered to regard the researchers as the “closest humans to the problem”.

Buckmaster then told OpenAI that if it released its result in the way proposed he would go public with what happened, to which OpenAI allegedly replied ominously: “Why would you ruin your career?” and “If you don’t want me to be nice, then I don’t have to be nice.” (Harris 2026)

On the same day as the two researchers, OpenAI also published its claim to have independently found the solution to the problem: “We’re sharing a solution to the Navier–Stokes existence and smoothness problem […] This proof, produced by an internal OpenAI system [..]”.

What on earth is going on?

Buckmaster notes how similar OpenAI’s solution appears to be to the two researcher’s own work, which, as he also notes, is “not the direction one arrives at in a few days by giving a model the problem statement.” (Buckmaster 2026)

In their own press release, OpenAI only acknowledges the two mathematicians achievement as “concurrent work”, it however admits that their “effort began on September 1st after hearing a rumor which we later realized was related to [Buckmaster and Alpöge]” (OpenAI Press Release).

But why? Why did OpenAI start throwing millions of dollars at a problem just about the time the two researchers were finishing up their work?

And there is a more severe suspicion: Did OpenAI monitor or trained its models on user data from private OpenAI Codex developer sessions to beat the researchers to a historic breakthrough? This would explain why OpenAI was able (as they claim) to solve the problem so quickly, why they solved it at the same time as the researchers, and why the approach to the solution looks so similar.

But wait, in their press release, OpenAI further states that “While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models” (OpenAI Press Release).

“We cannot rule [it] out”

I had to stop at this for a moment. What does this mean? Calling it “unlikely” means there is the possibility they have used material from a private Codex session (“their usage of our product”) without consent, or otherwise got hold of the two academics’ working papers? That alone would be alarming.

But why can’t OpenAI rule it out? Do they not know if they scan such repositories for training data? Are they not keeping a log of what training data is being used? If they keep a log, then this could be easily verified. If they do not, it shows either complete lack of consideration for effort that claims to be a scientific breakthrough, or a deliberate decision for deniability after the fact. Is OpenAI taking everyone for fools?

And if there was any doubt, why did OpenAI made their findings public?

Enough of that game

Over the last week, OpenAI has accompanied the release of its GPT-6 model (“Astra”) with wild claims. CEO Sam Altman has even claimed “AGI has been achieved”. As we have pointed out before, OpenAI likes to shroud the capabilities of its models in mystery and wonder.

However, the recent events around OpenAI and its AI agent’s hacking of Hugging Face And a German wiki site have damaged the company’s reputation, and assigning scientific breakthroughs to its newest models would certainly be a public relations stunt for the company.

OpenAI does not deserve the benefit of the doubt here. They need to deliver proof of its model’s capabilities, and extraordinary claims require extraordinary evidence. The burden is not on outside researchers to prove OpenAI’s claims wrong.

Hand waving away concerns with the argument “we do not know what our model was trained on” does not cut it. It should never be sufficient, but even less so when claiming a scientific breakthrough.

The EU AI Act requires GPAI providers to disclose their training sources to the regulator. Unfortunately, the act only requires companies to describe the source material in general terms, and does not mandate them to keep a detailed record.

This may have to change. Otherwise, what comes next? Will AI companies file patents hours before the inventor gets to do it, based on the researchers drafts? Or publish a novel with the same plot line as a human author days before them?

Conclusion

What do we learn from this episode, however it eventually plays out?

• Researchers should not use OpenAI’s tools if you are concerned they nick your work. Because they may, knowingly or not. Allegedly they do not know.
• None of OpenAI’s claims about “scientific achievements” claims are credible from here on, or should be taken seriously if there is the possibility that they are scanning researchers’ work in progress is what they do.

As Buckmaster notes, of course neither he not anybody can prove that OpenAI took their papers and trained their model. That is because OpenAI offers no transparency.

What did OpenAI know, and when did they know it?

The best scenario is that OpenAI is careless and sloppy. The worst case scenario is an egregious case of scientific misconduct. And OpenAI can not just put blame on its AI agents. There are people involved.