Bold editorial title card on cream. Sage letterspaced kicker reads NEWS DAY . FACULTY above a charcoal serif headline reading Nobody Asked For A Detector, with a sage italic subtitle reading They asked for the file. A sage seedling mark sits top left and a gold brushstroke arc curves over a sage circle at the right edge.
| |

Mathematics Settled Its AI Dispute With Receipts. Your Syllabus Asks for a Promise.

ByDean Le Blanc, Founder Harvest Kernel

Nobody ran the ninety year proof through a detector. They asked for the file.

On Tuesday, September 8, OpenAI published what it says is a resolution of the Navier Stokes existence and smoothness problem, one of the seven Millennium Prize Problems named by the Clay Mathematics Institute in 2000. The company says the proof came out of an internal system working through a group of coordinating agents, and that the group which produced the result ran on the order of ten thousand concurrent agents at once.

Here is the part worth your attention. In the twenty four hours that followed, the mathematics community argued fiercely about that announcement. It did not argue about whether a machine had written it. That question was settled before the conversation started, by the authors, on purpose, because both sides had already shipped the thing that settles it.

The first thing they shipped was a file a machine can check

OpenAI did not publish a proof and ask to be believed. It published, in its own words, “both a writeup of the proof and a formalization in Lean.” Lean is a proof assistant. A formalization in Lean is a document a computer can walk through line by line and either accept or reject, without knowing or caring who or what produced it.

The mathematicians on the other side of this week’s dispute did the same thing. Tristan Buckmaster of New York University and Levent Alpöge released results on a related fluid problem on September 7, the day before OpenAI’s announcement, and Buckmaster’s public statement records the date their own machine check finished: “we verified it on Lean on August 22nd.”

Read that again, because it is the whole lesson. Two groups who fundamentally disagree with each other about credit, priority and conduct agree completely on one thing: the argument does not start until the checkable artifact is on the table. Neither side asked the other to prove a human did the typing.

The second thing they shipped was a record

Once the artifact exists, an interesting thing happens to the disagreement. It stops being about the work and becomes about the trail.

Buckmaster’s statement is four pages and it is almost entirely a timeline. September 3, he emails a mathematician at OpenAI. September 4, he is asked to meet that day and declines. September 6, he is asked to meet “at any point today,” and two calls happen that afternoon. He quotes his own email in full rather than summarizing it, and he says why: “I would rather the full text be read rather than my summary of it.”

The central question he raises is a provenance question, and he states it as a question rather than an accusation: “I asked whether the model had been trained on, or had access to, our sessions in Codex, into which we had been putting all our drafts for the whole of this project.”

That is not a question a detector can answer. It is not a question any tool can answer. It is answerable only from a record of what was made, when, and from what. When the output itself can no longer tell you where it came from, the record is the only thing left that can.

The third thing they shipped was a list of what they were not claiming

The most professionally impressive paragraph of the week is the one where Buckmaster stops and marks the edges of his own knowledge:

“I would like to be clear about what I am not claiming. I have not seen OpenAI’s proof. I do not know what their model did, or how. I do not know whether our data was used. I am not accusing anyone of anything. I am stating what I was told, when, and what was proposed to me.”

Five sentences, and every one of them subtracts something. He is more credible after that paragraph than he was before it, and that is not an accident. In a week when the ceiling on what a machine can produce moved, the scarce and valuable thing was somebody drawing a clear line around what they personally knew.

Buckmaster also names, without being asked, whose idea the underlying program was. He credits Diego Córdoba and Luis Martínez Zoroa for the line of attack, says he and Alpöge took their work as a starting point, and writes that he believes Martínez Zoroa deserves a Fields Medal. He is in the middle of a priority dispute and he spends part of his statement giving credit away.

He also tells you, directly, that this is about your classroom

You do not have to extrapolate the teaching implication from a fluid dynamics story. The mathematician in the middle of it does it for you, in his own statement:

“the important thing is instead the significance that a mathematician and an LLM model can now do all this work in a month. The significance of this with respect to the way we train students, assign credit, referee, and decide what is worth one human life’s attention cannot be understated.”

Train students. Assign credit. Referee. That is a course design list, written by a working mathematician on the worst week of his professional year.

Want to do this on one of your own assignments this week? We built a free working session that takes one assignment you are already giving and adds the three things a promise cannot give. It takes about twenty minutes, it is done by hand, and you finish holding the rewritten assignment. Open The Checkable Step.

Your syllabus asks for a promise

Open your academic integrity language and read what it actually asks a student to do. In almost every version of it, the mechanism is a promise. The student affirms the work is their own. Some versions ask them to disclose AI use. A few ask for a reflection. All of them are a statement about the past, made by the person with the most to lose from telling the truth, and verified by nobody.

A promise was a reasonable instrument when the work carried its own evidence. A rushed essay looked rushed. A borrowed argument sat oddly in a student’s voice. You could read the artifact and infer the process, and the promise was a formality on top of an inference you had already made.

That inference is what stopped working, and no detector is going to give it back to you. The field that just lived through the sharpest version of this problem did not go looking for one. It asked for three things instead.

The three things, and what they look like in a course

One. An artifact somebody else can check without you in the room. Not a longer essay. A different kind of object. The dataset with the query that produced the table. The three sources with the sentence each one is actually supporting. The counterargument named and answered. In mathematics this week it was a Lean file. In your course it is whatever a colleague could pick up cold and evaluate on its own terms.

Two. A record of how the work was made. Dates, order, and what changed. Buckmaster’s timeline is doing exactly the work a process record does in a course: it makes a claim about the past checkable against something other than the author’s word. This is cheap to ask for and it is the single highest value line you can add to an assignment this term.

Three. A statement of what the student is not claiming. This is the one nobody assigns and it is the one that changes the most. A student who has to write down which parts they cannot vouch for has to know which parts those are. That is the entire skill. It is also the sentence that makes the rest of the submission believable, exactly as it did for Buckmaster.

None of the three requires a tool, a policy committee, or a detection subscription. All three are assignment design, and you can add them to one assignment this afternoon.

What this does not mean

It does not mean the OpenAI result is confirmed. A formalization in Lean establishes that a specific formal statement follows from specific formal assumptions. Whether that formal statement is the Millennium Prize problem as posed, and whether the surrounding argument holds, is what the mathematics community will now spend months on. OpenAI itself wrote that it does not intend to claim the Millennium Prize for the result. Buckmaster wrote that he has not seen the proof. Nobody in this story is claiming the matter is closed, and neither is this article.

It also does not mean process records solve academic dishonesty. A record can be fabricated. What a record changes is the cost and the shape of the dishonesty, and more importantly what it changes is the thing you are teaching. Asking a student to show the trail is asking them to work in a way that leaves one.

The one thing to take into Thursday

Pick the assignment you are least confident about. Not the one you suspect somebody of. The one where, if a colleague asked how you know the student did the thinking, your honest answer is that you are going on the writing.

Add one checkable artifact to it. Add one line asking for the dates. Add one line asking what they are not claiming.

That is a twenty minute edit and it moves the question you are asking from one nobody can answer to one anybody can.

The mathematicians did not get a better detector this week. They got a harder question and answered it in public with dated files. Your students are being asked for a signature.

Sources

When it is every assignment in every section. The Assessment Builder in the Faculty Toolkit writes assignments, rubrics, and tagged question banks against your own Course Record, so the checkable step is built into the assignment instead of bolted onto it afterward.

Still thinking it through. The Harvest Kernel Learning Community is free, and this week’s discussion thread is exactly this question.

Similar Posts