Judgment Is Not One Thing

In 1999, a researcher named Sugata Mitra cut a hole in a wall in a Delhi slum, mounted a computer behind it facing the street, wired it to the internet, and walked away. He left no instructions. Within hours, children who had never seen a PC and spoke little English were browsing, drawing, and teaching one another to use the mouse. Mitra ran the experiment again in village after village and kept getting the same result. He called it minimally invasive education. The rest of us called it surprising, and then, slowly, we stopped finding it surprising at all.

I keep returning to that image — the cluster of kids around a screen embedded in a wall — because it is the cleanest refutation I know of a comfortable assumption: that people need to be carefully prepared before they can use a powerful tool. They do not. They need the tool to be present, cheap, and consequential, and they will build the habits themselves. The preparation we design is mostly for our own comfort.

The history of technology adoption in the developing world is a history of this lesson repeating. Large parts of sub-Saharan Africa never built out copper landline networks; they went straight to mobile, and because there was no legacy to protect, adoption ran faster and innovation ran stranger. By 2007 Kenya had M-Pesa moving money over feature phones years before the rich world took mobile payments seriously. The infrastructure you skip is the infrastructure you are not held back by. Nicholas Negroponte's One Laptop Per Child, launched in the mid-2000s, got the economics half-right and the politics half-wrong, but its founding instinct — put the device in the child's hands and trust the curiosity — was correct in a way the program's critics never quite conceded.

I raise all of this because we are about to do it again, at a scale and speed that dwarfs every prior case, and the people who will be most affected are not the ones currently arguing about it. Inference is moving onto the device. Models are getting smaller, faster, and very nearly free to run. Within a few years the marginal cost of asking a capable model a question will round to zero, and it will happen on a phone in a village with no broadband, the same phone that leapfrogged the landline. The 99% who are not on LinkedIn debating prompt engineering, who are not enrolled in any program, who will never see the inside of a corporate clean room or a Socratic seminar — they are going to get AI anyway. The mouse is already in their hand.

So the question is not whether they will adopt it. They will. The question is what they will need to bring to it — and here the industry has converged, with suspicious unanimity, on a single answer.

I — THE LEDGER

On June 14th, Satya Nadella published a long note that most readers filed under the usual heading — AI, agents, enterprise — and most readers thereby missed the interesting part. His argument was not about technology. It was about economics, and specifically about what holds its value when intelligence becomes a commodity.

Nadella splits the value of a firm into two ledgers. There is token capital — the AI capability a company builds and owns — and there is human capital, which he defines as the knowledge, judgment, relationships, ingenuity, and pattern recognition of its people. His central, counterintuitive claim is that the second does not depreciate as the first grows. It appreciates. "Without human direction," he writes, "you have compute running in circles." The accompanying mechanism is what he calls a learning loop: the workflows, corrections, and accumulated business judgment that a company feeds back into its systems so they compound over time. The model, in his telling, is not the moat. The loop on top of it is.

But the note's actual title is not about any single firm. It is that a frontier without an ecosystem is not stable, and its deepest claim is political-economic. Nadella's real worry is concentration — that if a handful of models capture all the returns while every industry's knowledge is commoditized out from under it, the result is the same hollowing-out he watched globalization inflict on the industrial economies, where the aggregate numbers looked fine right up until the displacement became a political fact. His prescription is distribution: value has to flow broadly, across every company and industry and country, or the arrangement does not hold. Hold that thought. It is a claim about firms, and it has a twin one level down.

The line in the note that has stayed with me is smaller and sharper: "you can never offload your learning." You can hand a task to a machine — you can hand it an entire job — but the developmental residue of having done the work, the calibration that comes from having been wrong and noticed, does not transfer. It stays in the person, or it is lost.

Figure 1 — Reading the ledger. We are not contradicting the ledger; we are reading where its weight will fall. Of the five elements, Knowledge is the one inference makes cheap — so PJRC substitutes Creativity for it, formalizing a thread ("ingenuity") Nadella had already written down.

This is where GRID42 makes its one deliberate departure. If you compress Nadella's five elements into a working taxonomy of four — which is what any operating framework eventually has to do — the tempting move is to keep Knowledge as a pillar. We do the opposite. Our framework is PJRC: Pattern, Judgment, Relationship, Creativity. Where a faithful compression of the ledger might land on Knowledge as the fourth axis, we substitute Creativity, and the substitution is the entire argument.

II — THE INVERSION

The case for dropping Knowledge as a durable human axis is not that knowledge stops mattering. It is that the kind of knowledge a person can be tested on — recall, lookup, the citation of a regulation, the recitation of a procedure — is precisely the kind that cheap inference reproduces on demand. When the marginal cost of retrieving a fact approaches zero, the scarcity, and therefore the value, migrates somewhere else. It migrates to compiled knowledge: domain expertise that has been integrated, structured, and made usable under pressure. It migrates to systems thinking, the ability to see how parts interact rather than merely to know the parts. And it migrates to the judgment required to deploy any of it in a particular, messy, one-off situation.

Nadella concedes the mechanism himself, and it is worth pausing on who is conceding it: his own note warns that AI will steadily absorb the expertise of people and organizations and commoditize it. That is the inversion stated by the party with the most to lose from overstating it. If the man selling token capital tells you that expertise commoditizes, the only open question is what you stand on once it does.

I think of this as the abundance inversion, and it is the same shape as every leapfrog in the opening of this essay. The landline did not become valuable as it became cheap; it became irrelevant, and value moved up the stack to what mobile enabled. Raw recall is the landline of the cognitive economy. We are about to skip it.

Figure 2 — The abundance inversion. The curves are illustrative, not measured — but the shape is the claim. The scarce, defensible layer moves up as inference goes free. What you can be quizzed on depreciates; what you can only build through experience appreciates.

If that is right, then teaching the 99% to compete with the model on recall is malpractice. We would be training people for the landline. The defensible thing to teach is the layer the model does not own: the integration, the framing, the call. Which brings us to the word everyone uses for that layer, and why it is failing us.

III — THE TRAP

Walk through any vendor deck, any policy memo, any conference keynote on the future of work, and you will hit the same load-bearing sentence: AI does the execution, humans provide the judgment. Keep a human in the loop. The human brings judgment. It is reassuring, it is nearly universal, and it is — in the precise sense — empty, because "judgment" is doing the work of at least five different cognitive skills that do not develop together, are not measured the same way, and do not transfer from one to another.

This is not pedantry; it is the difference between a curriculum and a slogan. When I went looking at how the field actually defines the term — across the AI labs, the economists, and the much older and more sober human-factors literature — the single word fractured into five.

Figure 3 — One word, five skills. "Keep a human in the loop" quietly assumes the fifth box — calibration — and assumes people are good at it. The evidence says they are not, unless it is built deliberately. A program that promises "judgment" without saying which of these five, and how it is trained, is selling the slogan.

There is evaluative judgment — the taste to recognize whether an output is good, which is what the lab founders mean when they talk about taste surviving. There is directive judgment — deciding what is worth doing at all, framing the problem, which the economists from David Autor to Ajay Agrawal locate at the center of value. There is expert judgment — reading the one-off, high-stakes, no-single-right-answer case, the thing a senior practitioner does and cannot fully explain. There is oversight and calibration judgment — the meta-skill of knowing when to trust the machine and when to override it. And there is ethical and accountability judgment — locating the line that cannot be crossed and the responsibility that cannot be delegated.

The reason this matters is that the human-in-the-loop promise rests almost entirely on the fourth one, calibration, and the evidence on calibration is brutal. Lisanne Bainbridge warned in 1983, in a paper the automation world still quotes, that taking the easy parts of a task away from a human makes the hard residual parts harder, not easier — and that humans are worst at exactly the vigilant monitoring we keep assigning them. Forty years later Ben Green surveyed dozens of policies requiring human oversight of government algorithms and found that people could not actually perform the oversight functions the policies prescribed; the human in the loop functioned mainly as a moral crumple zone, a place to assign blame, not a real check. Daniel Kahneman and Gary Klein, in their rare published agreement, set the condition precisely: intuitive expertise is real, but only develops in a high-validity environment with enough feedback-rich repetitions to learn its regularities. Absent those conditions, what feels like judgment is just confidence.

The good is the leapfrog: capability delivered to people who were never going to get it through the old institutions. The bad is deskilling — the quiet atrophy of the very judgment we claim to be elevating, as people stop doing the reps that built it. The ugly is the false credential and the oversight theater: certificates that attest to "judgment" no one measured, humans placed in loops they cannot actually control, and a widening gap between those who already had domain expertise — for whom AI is rocket fuel — and those who did not, for whom it is a crutch that prevents them from ever building it. You do not get the good while pretending the bad and the ugly away. You get it by being specific about which judgment you mean and honest about how it is built.

IV — OUTSIDE THE CLEAN ROOM

Here is the hard fact the slogan obscures: every one of those five judgments, where it is real, was historically built the same way — by repeated exposure to consequential cases with feedback from someone who already had it. Apprenticeship. The case method. Clinical rounds. The cognitive apprenticeship that makes a master's tacit reasoning visible long enough for a novice to copy it. None of these scale. They are expensive, slow, and rationed by enrollment, which is precisely why judgment has always been the possession of the credentialed few. The bottleneck was never the knowledge. It was the coached repetitions.

And that bottleneck is exactly what cheap inference dissolves. The thing that scales worst in the old world — a patient, knowledgeable counterpart who will run you through case after case and tell you where you were wrong — is the thing a sufficiently well-designed system can now deliver to anyone, on the leapfrogged phone, at the cost of a query that rounds to zero. This is the opportunity hiding inside the disruption. We can finally take the apprenticeship out of the clean room.

But only if we build it deliberately, against the evidence, which is what GRID42 is. The three products are not a suite in the marketing sense. They are the three moves of a single loop — the same learning loop Nadella described for the enterprise, run instead for a person. He called it a cognitive loop, and then a hill-climbing machine, and his one insistence was that unlike most assets it compounds: every improved pass generates a better training signal, which sharpens the next. The same is true at the scale of one mind. Every case a person works produces a truer reading of where their judgment actually is, which sets up the next case. The compounding Nadella reserves for the firm is available to the individual — but only if someone builds the loop for them, because no one builds it for themselves.

V — THE LOOP

Figure 4 — The learning loop, for a person. The wall between Cubelet (knowledge, studied in the Library) and Case (judgment, practiced in the Lab) is the whole discipline of the build: it refuses the conflation the rest of the field makes.

Scaffold is the diagnosis: where do you actually sit on Pattern, Judgment, Relationship, and Creativity? The discipline here, and the place most assessments cheat, is the J. Kahneman and Klein's other finding is that subjective experience is a poor guide to the accuracy of one's own judgment — people cannot feel the difference between their good calls and their lucky ones. So a diagnostic that asks you how good your judgment is will measure your confidence, not your judgment. Scaffold has to probe the J by performance, against cases with known answers, or it is measuring the wrong thing.

Pursuit is where the gap closes, and it is built on a distinction we enforce at the level of vocabulary. A Cubelet is a unit of knowledge — studied, in the Library. A Case is a unit of judgment — practiced, in the Lab, against a fictional organization with embedded ground truth. They are never the same artifact and never called the same thing, because the moment you let knowledge consumption masquerade as judgment practice you have rebuilt the trap. The Cubelet feeds the compiled-knowledge layer the abundance inversion says is scarce; the Case manufactures the coached repetitions that judgment, and only reps, can produce. The Case is the thing that does not scale in the old world and now does.

And because judgment is five things, the Cases are not one thing either. A Determination case trains evaluative judgment against a clean verdict. An Override case trains calibration — scored not on a single answer but on whether you over-trust or under-trust across many — which is the one the human-in-the-loop world most needs and least measures. A Disposition case trains directive framing under constraint. A Diagnostic case trains expert reading of the messy whole. A Boundary case trains the accountability call. Each needs its own structure of ground truth, which is the unglamorous engineering reality behind the tidy slogan.

There is a darker reading of Nadella's loop, and it is worth saying plainly, because it is the reason the third product has to exist. When he writes that employees will see their judgment folded into systems that make it replicable and scalable, that is amplification described from the firm's side of the table. From the worker's side the same sentence can read as extraction: the veteran's hard-won judgment is decanted into the company's loop, the loop compounds, and the veteran becomes the part that is now replicable — which is to say, replaceable. This is precisely the deskilling the research warns about, sitting quietly inside the optimistic line. The only defense is ownership. If the judgment a person builds is attested to that person, in a credential they carry out the door, it becomes an asset the firm's loop cannot silently absorb.

Passport is the proof, and it is where the temptation to overclaim is strongest and most dangerous. The honest credential certifies only what is falsifiable: performance on freshly generated cases the holder has not seen, scoped explicitly to a domain — judgment in CMMC Level 2 assessment, not "judgment" — with the calibration error stated as a number and the methodology exposed so a relying party can check it rather than trust the word. What the Passport must never certify is Cubelet completion dressed as judgment, or some portable, general faculty the evidence says does not exist. A credential that survives that discipline is rare precisely because almost no one imposes it. It is also the only kind worth carrying through the wall, into a labor market full of certificates that attest to nothing.

VI — THE WAGER

I do not know what the settled norms of an AI-saturated working life will look like, and I am suspicious of anyone who claims to. We are at the stage the children at the wall were at in their first hour — touching the thing, finding out what it does, before any of the habits have formed. The habits will form. The norms will be written, mostly by people who are not in the current conversation, on devices that leapfrogged the infrastructure we think is load-bearing.

What I am willing to wager is narrower. It is that the comfortable story — humans keep some single, mystical faculty called judgment that the machine can never touch — is both false and lazy, and that its laziness is the danger. Judgment is not one thing. It is five things, built one way, by reps in a high-validity environment with honest feedback, and that way has always been rationed to the few who could afford the apprenticeship. The opportunity in front of us is not to protect that scarcity behind a slogan. It is to take cheap inference and use it to do the one thing the old institutions never could: manufacture the coached repetitions at the scale of the 99%, name precisely which judgment each one builds, and prove only what we have actually measured.

Which returns us to where Nadella started. His warning was that a frontier without an ecosystem is not stable — that value concentrated in a few hands, with everyone else's knowledge commoditized out from under them, is an arrangement the political economy eventually refuses. He meant it about firms and models. It is at least as true about people. A frontier without a workforce that can build judgment — not merely consume it, not be quizzed on it, but build it — is not stable either. The few who already had the apprenticeship compound their advantage; the rest are commoditized exactly as he fears industries will be, and the instability he warns about simply arrives through the labor market instead of the cap table. The ecosystem he wants for models is the one we have to build for judgment, and it does not build itself.

You can offload the task. You can offload the job. You cannot offload the learning — which means the learning is the asset, and building it, deliberately and at scale and outside the clean room, is the work. The mouse is already in their hand. The only question is whether we put anything worth practicing on the other side of the wall.

The Scaffold — a GRID42 publication. Framework: Pattern · Judgment · Relationship · Creativity. Products: Scaffold (know the gap) · Pursuit (close it) · Passport (prove the judgment).

Sources referenced: S. Nadella, note of June 14, 2026; L. Bainbridge, "Ironies of Automation" (1983); B. Green (2022); D. Kahneman & G. Klein (2009); D. Autor; A. Agrawal. Figures 1–4 are illustrative.