On the first of September, 2026, somebody at OpenAI heard a rumour. Two mathematicians, Tristan Buckmaster at NYU and Levent Alpöge, who works at Anthropic, were close to the Navier-Stokes existence and smoothness problem, a Millennium Prize question that had stood since 2000. The company pointed an unreleased model at it. Eighty-eight hours later it had a proof, and on September 8 it said so. Buckmaster asked whether his unpublished work had been fed to the machine. He says he never got an answer.
Three days after that, 25 Fields Medallists signed a declaration. On October 6 the company posted 722 manuscripts to GitHub in one evening. The next morning a group calling itself the Association for Human Mathematics asked its members to stop working with OpenAI and called the release "not a demonstration of scholarship, but a demonstration of power."
I have been in publishing and software since 1991 and I have never watched a profession go through the whole cycle in 32 days. Sarah Wilson's newsletter landed this afternoon telling her readers to trust their tears. Fair enough. I wanted to know what the tears predict, because I have shipped games, a stock engine and books with the same machines and taken the heat from three crafts for it. The script is the same each time. Only the page number changes.
The script, as mathematics ran it
| Date | What happened | Stage |
|---|---|---|
| Oct 2025 | OpenAI staff say GPT-5 "solved" ten Erdős problems. The man who runs the Erdős site points out it found old papers he had not read. Posts deleted. | 1 Denial |
| Sep 8, 2026 | Navier-Stokes claim. Buckmaster and Alpöge say the method was theirs and ask whether their data was used. | 2 Credit fight |
| Sep 11 | A Severe Misalignment of AI in Mathematics, 25 Fields Medallists: solving problems "is only a tool and proxy" for understanding. | 3 Declaration |
| Sep 21 | OpenAI names an advisory group of nine mathematicians, unpaid, no power. Scientific American reports the proof solved the easier variant. | 4 Governance |
| Oct 6 | 722 manuscripts, 372 results, one prompt. Three withdrawn within days over a sign error. | 5 The dump |
| Oct 7 | "Mathematicians did not ask for this work to be done." Stop working with OpenAI. | 6 Boycott |
| Oct 9 | Terence Tao at Caltech: "Math 2.0." Proofs are abundant now; chasing them "is now actively harmful to the long-term health of mathematics." | 7 Redefinition |
Stage one bought the field eleven months. Every other profession on this page is somewhere inside its own eleven months.
What the two camps are actually fighting about
Peter Scholze told IEEE Spectrum the labs are "just solving these difficult mathematical problems as benchmarks, as some kind of PR stunt." A PhD student in Amsterdam told the same reporter she is leaving the field "because I don't want to be in a career where you're verifying LLM output." On the other side, Daniel Litt at Toronto called the release "great for mathematics," and OpenAI says it wants "to push the frontier of human knowledge." Andrew Sutherland at MIT said the only thing an engineer would say: "We should ask for receipts."
The fight underneath
Nobody disputes that the machine can write proofs. The fight is over who gets to say what the job was for. For sixty years the profession measured itself by solved problems because solved problems were scarce and hard to fake. The labs took the yardstick literally and built a machine that maxes it out. Every profession below has a yardstick like that.
Four defences decide who goes next
Mathematics did not fall first because it is easy for machines. It fell first because nothing stood between the machine and the work. Four things can:
- A stamp. A licensed human must sign and carry the blame. Engineers, physicians, lawyers, auditors. Mathematicians sign nothing.
- The customer. The practitioner owns the relationship, not an employer who can buy software over their head.
- A union. Anger becomes contract language inside eighteen months. Without one it becomes open letters.
- A closed archive. The profession's back catalogue was not the training fuel. Where it was, a wage fight becomes a consent fight, which lasts longer.
One more dial sets what the fight is about, not how hard it is: whether the deliverable is the answer (a brief, a bridge) or the process (a proof, a degree). Answer professions fight over price. Process professions fight over meaning.
Mathematics had every dial on zero. So it went first.
The four defences, measured for mathematics in September 2026
Where each profession is, and what I expect next
Dots are defences held, out of four: stamp, customer, union, closed archive. A half dot is a partial hold. The dates are where I will be wrong in public.
| Profession | Defences | Where it is today | What happens next, and when |
|---|---|---|---|
| Mathematics | 0 | 7 The full script in 32 days. Three of 722 papers already withdrawn. | Journals demand receipts (model, prompt, compute) by mid-2027. Pure-math enrolment drops; the boycott fails as a boycott and wins as a norm. |
| Theoretical physics | 0 | 1 Tao says AI already saves more time than it wastes here. Anthropic ran an unreleased model at the Riemann hypothesis in August. | Same dials as math, same script, one year behind. A lab claims a big theoretical result by late 2027 and a physicists' letter follows inside a month. |
| Software | 0 | 5 curl ended its bug bounty in January after AI reports swamped real ones; Ghostty and tldraw shut the door on outside pull requests. Workers aged 22 to 25 in exposed jobs sit 19% below trend. | Already past the math stage, with no famous names to make it a story. The junior pipeline stays broken through 2028. No boycott: the people drowning in AI code are the ones writing it. |
| Translation, voice, illustration | 0 | 6 Korean voice actors report incomes halved in a year or two. A Rest of World tracker counts over 100 creative-worker movements in 25 countries. | The clearest wage loss on the list and the loudest consent fight. A standard voice-rights contract in a major market by 2027. The middle of the market goes; the certified and the famous stay. |
| Journalism | ½ | 3 Pew: 59% of the public and 60% of AI experts expect fewer jobs. The fight is over the archive, priced by a $1.5 billion authors' settlement. | Licensing invoices, not bans. Every major paper has an AI licence by 2027 and keeps shedding reporters regardless. |
| Actors and writers | 1 | 6 SAG-AFTRA on the AI "actress" Tilly Norwood: not an actor, a program "trained on performers' work without permission." WGA signed a four-year deal with AI limits in April. | The one unstamped profession that won, because it struck. Holds to 2030. The fight moves to non-union production and overseas dubbing, where it is already lost. |
| Teaching (university) | 1½ | 3 Harvard faculty, September 2026: 64% say AI has harmed their courses, up from 42%; 24% ban it. | The fight is assessment, not the tool. Oral exams and proctored essays are the default by 2027; the take-home paper is gone by 2028. |
| Nursing | 2 | 6 Twelve Montefiore nurses laid off for software in July, grievance filed. Nurses protested Palantir in eight cities on August 27. | The most organised fight in health care: a union, and an employer who owns the patient. An AI-staffing clause in every big nursing contract renewed in 2027, and one strike with AI scheduling named. |
| Truck drivers | 2 | 6 Teamsters sued California on August 5 over driverless heavy-truck rules, "an existential threat" to hundreds of thousands of jobs. | A job fight, not a craft fight. Human-operator bills pass in union states and fail in Texas. The experts who expect driver losses are right, slowly. |
| Law | 2 | 4 Sanctions for invented citations: $5,000 in 2023, $110,000 in 2026. 69% of legal professionals use general AI; 9% of firms enforce a policy. | Lawyers never fight the tool; courts fight the lawyers. AI disclosure becomes a line in the signature block by 2027, and the first-year class at big firms is a third smaller by 2028. |
| Accounting and audit | 2 | 5 KPMG cut about a tenth of its US audit partners this year, citing AI audit tools (one trade report). Entry-level hiring down three years running. | Quiet from above, collapse from below. The junior work that trained partners is what the machine eats. A regulator's problem by 2028. |
| Engineering (licensed) | 2½ | 0 No organised pushback anywhere I looked. Liability stays with the engineer who stamps. | Nothing until a stamped, AI-generated design fails in public, somewhere between 2027 and 2029. Then rules about disclosing AI in stamped work. Until then it hollows out like accounting. |
| Medicine (physicians) | 2½ | 4 AMA, 1,692 physicians, early 2026: 81% use AI, up from 38% in 2023; 88% fear skill loss; nearly half strongly oppose patients reading their own scans with AI. | No boycott, ever. The fight is the patient who arrives with an AI second opinion, peaking in 2027. A malpractice case where the machine was right and the doctor overrode it, before the end of 2027. |
| Therapy and counselling | 3 | 4 Illinois banned AI therapy without a licensed professional in August 2025, after Nevada and Utah. Fines to $10,000. | The stamp arrived by legislation before the machine did. Most states covered by 2027, enforced against companies. Clients keep using the chatbot at 2 a.m. anyway. |
Engineering, medicine and law
Engineering is the quietest profession on the list, and that is the warning. The licensed engineer stamps the drawing and carries the blame, so the tool gets absorbed the way CAD was in the 1990s: seniors check the output, juniors who used to produce it are not hired, and everyone at the top reports record productivity. The stamp protects the stamp-holder, not the profession. In fifteen years it produces seniors who never did the calculation by hand and cannot check the machine. The backlash starts the day one of them signs something that falls down.
Medicine's fight is in the waiting room. I am 60 and on a longer list of drugs and supplements than any one physician has time to hold in their head. The AI I use remembers all of it and reads a new symptom against the whole list in ten seconds. My doctors cannot; they have eleven minutes and four hundred other patients. The AMA's own survey says 81% of physicians now use AI and nearly half strongly oppose patients using it on their own results. Both are true at once: the profession adopted the tool and wants to keep it out of the patient's hands. That is the stamp doing its job. There will be no declaration, because 81% of the signatories would be denouncing the thing they use. There will be a malpractice case where the machine was right.
Law is the easiest call, because its product is the answer. The lawyer signs, owns the client, and is paid for the brief, not the understanding. So the profession absorbs the machine and polices the output, and the courts do the policing: the fine for an invented citation went from $5,000 to $110,000 in three years, and the same report says the filings have not stopped. Sixty-nine percent use the tool and nine percent of firms have a rule about it. That gap is the entire backlash. Document review was the apprenticeship, and document review is gone.
Go already ran this script
In November 2019 Lee Sedol, the only human ever to win a game against AlphaGo, retired from professional Go. "Even if I become the number one, there is an entity that cannot be defeated." Five years later: losing to the machine "meant my entire world was collapsing. I could no longer enjoy the game." Go did not die. Its professionals train against engines and are stronger than any generation before them. What died was the thing he was in it for. Tao's "Math 2.0" is the Go answer, written before the retirement wave instead of after it.
What I'm not sure about
- Whether the 722 results hold. Three fell in a week. If a large share fail, stage five was a bluff and the script slows everywhere.
- Who the Association for Human Mathematics speaks for. Its statement is signed by a working group, not by names.
- Engineering. My prediction needs a failure. If none comes, the profession hollows out with no fight at all, which is worse and invisible.
The size of the backlash does not measure how good the machine is at the job. It measures how few defences the profession held when the machine arrived. If you are in one of these fields, do not ask whether the machine can do your work. Count your dots.
Sources
- The declaration, Sep 11, 2026. Tao's blog · AHM statement. ahmath.org · Tao's "Math 2.0" and disclosures. tao-web
- The Oct 6 release, Sutherland, Litt. Scientific American, Oct 6 · withdrawals, OpenAI's line. Fortune, Oct 7
- Scholze, Robertson, the Heidelberg reaction. IEEE Spectrum, Oct 5 · the Sep 1 rumour and 88 hours. The Conversation, Sep 9 · the easier variant. Scientific American, Sep 21 · advisory group. OpenAI, Sep 21
- The Erdős episode. TechCrunch, Oct 2025 · Sarah Wilson. This is Precious, Oct 14 · Lee Sedol. CNN, 2019
- AMA 2026 survey, 1,692 physicians. press release · report (PDF)
- Montefiore nurses. Bronx Times · Palantir protests. KQED, Aug 28 · union contracts. Marketplace, Jun 24
- Sanctions, the 69% / 9% survey. Newsweek, Sep 24 · $5,000 to $110,000. Law.com, Sep 30
- Young workers 19% below trend. Stanford Digital Economy Lab, Aug 12 · curl, Ghostty, tldraw. The New Stack; codenote
- KPMG partners, single trade report. TheStreet · engineering liability, commentary. AI Standard of Care
- Harvard faculty survey. The Crimson, Sep 14 · voice actors. Rest of World, Apr 15; Korea Times, Apr 24 · translators. CNN, Jan 23
- SAG-AFTRA. Variety · WGA deal. NPR via WLRN · Illinois. Consumer Federation of America · Teamsters. E&E News, Aug 6 · public vs experts. Pew, Apr 2025