A Working Question for the Last Good Moment

Your author on the crossroads: Someone has started fiddling with the lighting.

This last spring, Claude and I wrote A User’s Guide to the Last Good Moment. The idea of the LGM was simple enough. A scientist is five things—Scholar, Calculator, Referee, Logician, Architect—and AI had already mastered the first two, was well on its way to the third, and was starting to make unsettling noises about the fourth. The fifth, the Architect who decides what is worth wondering about, was still in the Human domain. We were living, we said, in the photograph we only recognize afterward: the one where everything is still fine.

Six months later, I am reporting back from inside the photograph. The kids are still young (at least in my heart) and the house is occasionally full. But we have come to a crossroads, and someone has started fiddling with the lighting.

Into the Arms of the LGM

The newest models from Anthropic and OpenAI have not so much nudged us toward the LGM as pushed us firmly into its arms, the way my dear friend Shaifali Puri would push me toward a stranger at her beloved Friday soirees with the words “you two will love each other.” Well, that remains to be seen. But we are pretty well acquainted. Tasks that I used to think of as mine—not just the drudge work but also some of the really interesting bits, now arrive three-quarters-done, and three-quarters-done rather well, by my buddies Fable and Astra.

But the LGM seems to be closing—as it must, I fear—in a smooth pincer movement. One jaw is closing on writing, another on formulating theorems and proving them. My own discipline, economic theory, sits more or less in the dead center of this two-pronged bite.

Writing and Proving

This past May, a short story called The Serpent in the Grove, by a Trinidadian writer named Jamir Nazir, was named the Caribbean regional winner of the Commonwealth Short Story Prize. Within days it had gone viral on social media. Critics claimed the story bore the “obvious markers” of AI. It contained, they said, too many “not x, but y” constructions and too many lists of three. Particular sentences were held up for inspection, among them the observation that a certain neighbor was big “in the way of women who never apologise to furniture.” (Go figure; or should I say go, Figure!) Someone ran the story through Pangram and reported that it flagged the piece at 100%-AI, adding, for good measure, “if you know you know.” Granta withdrew from its long-standing arrangement to publish the Commonwealth winners.

The Commonwealth Foundation did something interesting. It did not run a detector. It asked its writers for their drafts, their time-stamped documents and their notes, and it pronounced itself satisfied. Nazir, for his part, described writing six or seven drafts on a phone using speech-to-text, polishing each line before moving to the next because he could only see three or four lines at a time. In July, the story was named the overall winner. Social media remained unconvinced. The Foundation’s director-general made a pointed observation: when the machine’s default voice is the metropolitan one, the writer who does not fit the expected mould is the first to be suspected, and the more unfamiliar her brilliance, the readier the accusation.¹

I’m not sure if Nazir used AI, and neither, I suspect, are you (nor the prize committee, or Granta, or a detection company and several thousand strangers). The lights are flickering.

Now walk with me to the opposite end of the intellectual spectrum, where nobody is arguing about an “artfully placed semi-colon,” to use Hilary Mantel’s immortal phrase.

In August, Tristan Buckmaster (from NYU) and Levent Alpöge (from Anthropic, though working on his own time) made significant progress on the Navier-Stokes problem: one of the seven Millennium Prize problems. AI helped them, of course, in a loop that will sound familiar to anyone who uses these tools seriously: set up the problem, feed it in, read what comes out, feed it back. In early September, before they had finished, OpenAI announced that its own systems had produced a proof of blow-up for the Navier-Stokes equations. There followed a weekend of resentment, refusals, and recrimination which I will not recount here, partly because NYU employs both Tristan and me, so I suspect my kneejerk reaction will be biased in favor of solidarity, and partly because Anthropic employs Tristan’s co-author, and I have reasons to like Anthropic.²

What interests me is something Buckmaster said in the middle of it all. Maybe AI had proved something, but the main intellectual credit for the strategy of attack, he insisted, belonged to two Spanish mathematicians, Diego Córdoba and Luis Martínez-Zoroa. Undoubtedly, the AI experience had been exciting and frightening in equal measure. And then he asked the question I want to borrow: What’s the human part of it?³

Small Doubts and Big Doubts

The two stories have something in common: that nothing is settled. Nazir was cleared, but not to everyone’s satisfaction. And it’s unclear if OpenAI blew up the “correct” version of Navier-Stokes, or what belongs to Buckmaster and what belongs to the AIs that were used. As I said: it’s the damn lights.

What they do not have in common is the capability in dispute. The Nazir affair is about a Scribe: a capability that did not even make it onto our original list of five. It was left off because I believe a great human writer will still unmistakably stand out, machine or no machine. So I am personally ok with AI doing the mundane (i.e., the economist’s mundane) task of writing.⁴ In contrast, the Navier-Stokes and related affairs are about a Logician at full strength. One rather amiable jaw of the pincer worries at the shape of your sentences. The other quite sinister jaw buries into the proofs of your theorems. These are very different anxieties, and it muddies the waters to talk about “AI” as if they were one.

Economics in the Bite

As I said, academic disciplines like economic theory sit in the bite of the pincer. It’s a poky and uncomfortable place to sit. I am now reading new papers by students and colleagues that are obviously written with the help of Claude or Chat. You can tell, clear as day. The prose is suddenly clean and fluent. The paragraphs are sensibly load-bearing. The writing is earning its keep. (Oops!)

But “obviously written with the help of AI” is multidimensional, and I want to unpack it. I will do this specifically for economic theory: the branch of the discipline that uses models and formal reasoning to convey economic ideas or to generate predictions.⁵

First, there is the business of using AI to write a paper: to formulate the sentences, get the grammar right, and arrange the paragraphs in an order that a reader might wish to follow. This does not worry me one bit. Half the people in my profession cannot write properly, or at least they write in a manner so soporific that it would put a whirling dervish to sleep mid-whirl. Another forty percent write correctly, I suppose, but as if they are recording the minutes of a meeting nobody attended. Therefore, to the world I say: bring on the AI! Claude can string a sentence together better than ninety percent of my profession. The remaining ten percent will do ok, except—alas—they won’t stand out as nicely as they should.

In short, I’m not much moved by the Nazir-style worry that clean prose is itself suspicious. In literature the prose (or the poetry) is most of the work. The worry makes sense there. In economic theory the prose is largely superficial packaging. Nobody ever got tenure for a well-dressed lemma, and nobody should lose it for one either. And as a bonus: never again will I have to read a development economics paper with the title:

“Does Blah cause Blah?: Evidence from Blah.” (Or perhaps a Randomized Blah.)

Spare me!

Reasoning is another matter, and here comes the pincer with teeth. AI is now used not to dress the argument but to build it. This comes in two varieties:

(a) AI involvement in the proof of a well-posed theorem: the theorem has been conjectured and formulated, and the question is how to establish it.

(b) AI involvement in the building of the model itself: deciding what the agents are, what they know, what they want, and what they can do about it.

The first is increasingly common, I think. I have used it myself, and I will not pretend otherwise. I earlier mentioned (with Claude) a change-of-variables argument that sent me into a mildly ecstatic state of panic. More recently, Astra set off another bout of the shivers when it quite beautifully established a result on inefficiency that I had conjectured but could not prove. And (b) will become increasingly widespread. It’s a familiar give-and-take in model construction that one can profitably do with a top AI. See, for instance, the recent set of videos by Pietro Ortoleva and Fedor Sandomirskiy published in the Markus Academy—a series hosted by Markus Brunnermeier at Princeton—on using AI to iteratively construct economic models. Both (a) and (b) will accelerate at a remarkable pace in the years ahead. Maybe I should say months.

And, of course, the Prisoner’s Dilemma awaits. If we refuse to use these tools, others will not. Research, whatever else it is, is partly a signal. If our goal is to signal our prowess at model-building and theorem-proving to the wider community (it may not be the only goal, but for many younger scholars it is a large chunk of the whole enterprise), then prim refusal is not going to work. You will be the one runner in the marathon who declined the new shoes on principle, but no one will be looking at your old shoes. Trust me.

What Is to Be Done?

So, as V. I. Lenin famously asked, What Is to Be Done? (Vladimir Ilyich, it should be said, borrowed the title from a novel by Nikolai Chernyshevsky, which makes him an early and distinguished contributor to the problem of attribution.)

The question is not academic, or rather, it is only academic, which is worse. Our profession runs on evaluation. We evaluate our colleagues for tenure and promotion, our PhD students for their degrees and letters, our new recruits on the job market, and each other, endlessly, in referee reports. How do we carry out this grand activity while living in the steadily darkening penumbra of the Last Good Moment, with the lighting under dispute?

Let me pose the question as follows. Suppose—because we are only in the Last Good Moment, and not yet beyond it—that we human economic theorists are still the Architects. We have the research ideas. We choose the topics. We choose the questions. So far, so good. Those blueprints are still in our hands. Now come three areas in different shades of grey:

1. Who converted the question into a workable formal model? That is the give and take, the iteration, that Fedor and Pietro talk about in their Markus Academy video series.

2. Who conjectured the theorems within the working model—who looked at the model and said, I bet this is true?

3. Who proved the theorems in that working model?

To each of these questions, I submit to you that the answer is increasingly likely to be: AI. True, we are only at the end of September 2026, and we can still have a nice debate about these. Enjoy it; that debate will soon be over.

AI, AI, AI, to paraphrase a Spanish lament. The Architect still holds the pencil. But the model, the conjecture and the proof—the three things we used to point to when we said this person is good—are drifting, one by one, into the machine. And unlike the Commonwealth Foundation, we cannot simply ask for the drafts. Those were written by AI too.

Which brings me to the question this essay has been circling. I’d like to say, like a tiger circling its prey, but I guess it’s more like a nerdy referee circling a paper they have already decided to reject. Buckmaster asked it of fluid dynamics, but I want to ask it of a discipline where the fluids are considerably less turbulent and the theorems considerably easier:

When a young economic theorist hands us a paper in the Last Good Moment, and the model may be the machine’s, and the conjecture may be the machine’s, and the proof may be the machine’s—

What’s the human part of it?

Maybe it all comes down to the original idea, the original insight? Maybe that's enough, on this new leveled playing field? That’s my working question as I ponder (this iteration of) the Last Good Moment.

---

Notes

1. The Guardian, 1 July 2026. ↩

2. You guessed it: it built my buddy and partner in all sorts of intellectual pursuits, the irrepressible Claude! ↩

3. The New York Times, September 10, 2026. ↩

4. As I am ok with the middle ground of blog posts, op-eds and everyday thoughts, where most of the public conversation lives: who cares who wrote them? Claude and I said our piece on that last spring. If the ideas are yours, the commas are a detail. And if there were no ideas to begin with, well, the machine has taken nothing that was there. TBH, this is more about having an educated reading public than having a Pangram-style AI-detector. I mean, look at the vocabulary of the New York Times’ Spelling Bee. But enough of pet peeves. ↩

5. I don’t think anyone will argue that the tremendous data-processing abilities that have arrived with agentic AI are anything other than an unambiguous positive. ↩

Note. With thanks to Claude, my co-author on the first instalment, who has been demoted to a footnote in this one and took it with a grace I found faintly suspicious. Thanks, too, to Parikshit Ghosh for painting humans into the ultimate corner—he will know what I mean.  In the interests of disclosure, and in the spirit of what I say above about writing: the sentences in this essay were drafted with Claude’s help from my notes. The ideas, and the grudges, are mine.