Complete English translation.

Preface · Form and Substance

A note on the basic concepts this essay uses.

Any deliverable — an article, a product, a proposal, a piece of code, a video — can be split into two parts: form and substance.

Form is the part of a deliverable that an external system can take over. It can be produced, copied, optimised, or scaled through tools, templates, conventions, processes, models, or division of labour.

Form is not synonymous with “low-grade” or “easy”. In some periods, form itself is extraordinarily hard. The clean handwriting of a scribe, the photographic likeness of a realist painter, the engineering craft of a programmer — each took fifteen years to master. But all of them are form, because one day external systems take them over.

Substance is the part the external system cannot bear on the subject’s behalf: redefining the problem, judging reality, taking responsibility for trade-offs, bearing the consequences, and the not-fully-transferable judgment formed through long experience. Substance is not “appearing distinctive”. It comes from one specific subject’s understanding of, judgment about, and answer for the real world.

Put differently: form answers “how is the thing made”; substance answers “why is it made this way, and who bears the consequences”.

Chinese has a settled legal and accounting principle that names this pair: 实质重于形式 — substance over form. In assessing the real value of a transaction or contract, you cannot stop at the surface arrangement; you have to penetrate the form and look at the underlying relationships and the real consequences.

This essay borrows that principle and places it in a larger frame.

History does not turn substance into form; it migrates form from the subject to the outside

There is a confusion to head off here.

Form and substance have been two different things from the start. What history changes is who completes the form. The definitions themselves have not moved.

In tool-poor eras, much of form could only be done by the person. Neat handwriting, accurate proportion, information retrieval, fluent prose, correct code — these are themselves form, but in a particular historical phase they had to live attached to specific people. Society identified and priced these form abilities as part of being that person. A scribe who wrote neatly, a painter who painted accurately, a researcher who held the literature in his head, an engineer who could write running code — in their respective eras, these were genuine scarcities.

When tools advance, the binding loosens. Each tool advance migrates form from the subject to an external system: print took over copying, photography took over likeness, industry took over precision, search took over retrieval. The next section walks through this line.

After form is taken over, it does not lose use value. What drops is form’s evidentiary power. It still works, still gets consumed, but it can no longer reliably testify that the subject behind it carries the corresponding substance.

AI is the largest migration on this line

AI is special because it does not just take over one kind of form. It takes over, in one stroke, a large number of form abilities that had still been residing inside the subject.

Writing a structurally complete, fluent long-form essay used to demand long training. After AI, this ability is rapidly externalised. A clean structure no longer reliably testifies that a writer who genuinely understands the problem stands behind it.

Writing a piece of well-architected, runnable code used to require an engineer who knew syntax, patterns, and debugging. After AI, the form-production of code is partially externalised. Working code no longer testifies, the way it used to, that the person who wrote it carries the corresponding engineering judgment.

Putting together a clear, well-organised proposal with diagrams used to require the training of a consultant, a designer, a product manager. After AI, that form is generated quickly. A proposal that looks complete no longer testifies that the people behind it actually understand the problem, the client, the constraints, or the consequences.

The abilities that an enormous number of professional workers built their livelihoods on over the last fifteen to twenty years have, since AI arrived, moved from “form that must be done by the subject” to “form an external system can do too”. This is a structural redrawing of the evidentiary relationship between form and substance, not just local replacement.

The downstream consequences run a layer deeper than “which jobs will AI replace”.

The question this essay starts from is here:

When form no longer reliably testifies to substance, where does substance come from, how does it get identified, how does it become value?

---

A History of Form Migration: Four Externalisations

This migration is supported by historical evidence. This section walks through four tool advances: print, photography, industrialisation, search. Each one shows the concrete shape of “form moving from the subject to an external system”, and the impact it had on the subjects of the time.

After these four examples, it becomes clearer that what AI is doing is not new. It is just the largest, fastest version on the same line.

Print (15th century): handwriting and copying externalised

Before printing, a finished book was a scarce signal. A bound, neatly written, accurately copied manuscript proved there was a trained scribe behind it, with enough time and enough resources. A scribe’s neat hand and accurate copying were, in the manuscript era, the central evidence on which he was priced as a subject.

Print externalised all of this. A press could output thousands of pages a day; a scribe, dozens. Two to three orders of magnitude. Between 1450 and 1500, book prices fell at least 65%. By Buringh and van Zanden’s estimate, Western Europe printed about 12.6 million books between 1454 and 1500, already more than the entire fifteenth-century European manuscript output.

By around 1525, humanists were starting to complain — there were too many books. In the Adagia, Erasmus wrote a line that reads almost word for word like today’s complaints about AI-generated content: “Is there anywhere on earth exempt from these swarms of new books?”

A scribe’s neat hand could no longer reliably testify to any single subject, because every printed page was neat. That part of form migrated from scribes to machines, and the scribe as a profession was devalued within a generation or two.

But substance was not destroyed. Literacy, reading, thinking, judgment all kept their value, and after print they became more valuable. Print made enormous quantities of text broadly available, and the people who were genuinely good at filtering, interpreting, and criticising became scarcer rather than less rare. What lost value was the externalised form. The substance that had been attached to it did not lose value.

Photography (19th century): representation externalised

On 19 August 1839, the daguerreotype process was unveiled at the French Academy of Sciences in Paris. The French government placed it in the public domain, and within months daguerreotype studios were spreading across Europe and North America. By the 1850s, commercial photography studios in Paris and other major cities were turning out portraits at a fraction — sometimes a hundredth — of the price of a painted portrait. Portrait photography went from novelty to middle-class consumer good in a few short years.

The people hit hardest were realist painters, who made their living “rendering reality faithfully”. A daguerreotype portrait took minutes; an oil portrait took weeks. Customers voted with their wallets. A line often attributed (probably apocryphally) to Paul Delaroche captures the shock with great precision: “From today, painting is dead.” He may not have actually said it, but the line is quoted again and again because it captures the real fear of the moment realist representation got taken over by a machine.

Realist painters faced exactly the same problem many professions face today: the core ability they had spent fifteen years training was taken over by an external system.

Painting did not die, but it had to redefine its value as something other than representation. Impressionism, post-Impressionism, Cubism, Abstract Expressionism — a full century of art movements were all, at bottom, painting’s answer to one question: once the camera takes over representation, what is painting’s substance?

Substance did not disappear. Representation, which had once been attached to the painter, migrated to the camera. The painter’s substance was redefined as the things the camera could not do for him: way of seeing, stance, style, judgment about what is worth painting at all.

Industrial production (19th-20th century): precision externalised

Industrialisation is form migration on a longer time scale. From British looms in the 19th century, to Ford’s assembly line at the start of the 20th, to Japanese lean production in the second half of the century — wave after wave migrated “evenness, precision, consistency, stable replication” from craftsmen to machine systems.

Before industrialisation, the fineness of an object often spoke for the craftsman’s experience, patience, and ability. The even stitching on a hand-made shoe testified that a trained master had made it.

After industrialisation, machines could mass-produce items more consistent than handwork. After Ford introduced the moving assembly line in 1913, chassis assembly time for the Model T fell from 12.5 hours to 93 minutes; the price dropped from $850 in 1908 to $260 by 1924. “Even stitching” no longer reliably testified to any craftsman’s substance, because the line’s stitching was more even than his.

But craft did not disappear entirely. After industrialisation, what survived turned toward what machines cannot do: special commissions, scarce materials, personal design, cultural memory, the beauty of irreproducible flaws. A pair of fully hand-made leather shoes is more expensive today than it was a hundred years ago. The precision may be the same. The substance attached to it has changed completely.

The internet and search (1990s-2010s): retrieval externalised

The most recent large-scale form migration is the search engine.

Before search, one of the core abilities of a knowledge worker was “knowing where to look”. Lawyers had to know the case books; doctors had to keep diagnostic criteria in their heads; journalists accumulated source networks; scholars commanded their bibliographies. These abilities — “information retrieval, materials organisation, basic fact-checking” — lived attached to specific people for years on end and were a substantial part of how they were priced as subjects.

Search engines externalised them. Google was founded in 1998; by the 2010s, an ordinary person could pull up a legal clause, a medical definition, a historical event, or a product spec in 30 seconds.

This migration unfolded slowly across the 2000s and was underestimated. It did not produce visible shock in two or three years the way AI has, but the cumulative effect was equally enormous. Large categories of “information-asymmetry work” — brokers, information clerks, junior research assistants, parts of paralegal work — slowly shrank over twenty years.

But again, substance did not disappear. Judging which materials matter, evaluating the quality of materials, applying materials in complex contexts, judging the real problem behind the materials — these abilities, in fact, became scarcer once search collapsed the information barrier. Being able to search is no longer worth much. Knowing what to search and how to use what you find is still worth a lot.

One shared trajectory

The four migrations follow the same trajectory:

1. A form ability lives attached to some professional group for a long time, treated by the market as visible evidence of their substance
2. A tool advance externalises this part of form
3. Form keeps its use value, but its evidentiary power for that group drops sharply
4. That professional group faces a collective devaluation
5. The ones who remain redefine value around substance the external system cannot do

In each migration, the people inside it felt they were facing an unprecedented crisis. In each one, substance did not disappear; substance had to be re-identified and re-priced.

AI is doing the same thing on an order-of-magnitude larger scale. It does not take over one form at a time; it takes over dozens at once.

The underlying logic does not change: form keeps its use value; it just no longer reliably testifies to the subject. That is the historical footnote to the preface’s central question: once form no longer reliably testifies to substance, where does substance come from, how does it get identified, how does it become value?

The next eight chapters unfold this question.

---

Chapter 1 · The Bell-Curve Society Is a Historical Exception

Before discussing AI’s effect on deliverables, the reader needs to be placed at the right historical position. One key fact is routinely overlooked: the “bell-curve society” we live in (most people in the middle, a few at each tail) has existed for less than 80 years in human history. It is an exception, not the norm.

For most of human history, the shape was a dumbbell. A small top, a large bottom, almost nothing in the middle. From the consolidation of European feudalism around 1000 CE to the start of the Industrial Revolution around 1760, this shape held steady for about 760 years; the agricultural population in most pre-industrial societies sat at 70-85%. In the United States in 1790, about 90% of the population lived on farms; even in 1860, 53% of the labour force was still in agriculture.

The Industrial Revolution started taking that structure apart, but not by turning the dumbbell directly into a bell curve. The bottom of the dumbbell shifted from agriculture to industry; the genuine swelling of the middle came much later — a transition that took about 180 years.

The forces that produced the bell curve in those 180 years were not a single variable. Late-19th to early-20th-century capitalist globalisation built mass production and division-of-labour systems; early-20th-century Fordism, electrification, and the assembly line drove industrial-good prices down to where ordinary workers could afford them; advertising, instalment payments, and mass consumer culture turned “middle income equals middle lifestyle” into a visible and reachable goal; post-war labour legislation, minimum wage, social security, and the GI Bill systematically pushed workers and veterans into the middle class. The bell-curve society is the product of a stack of forces, not the natural result of industrialisation marching forward.

The “middle class” in the modern sense — a society in which most people belong to that class — only began to form in the United States in the 1940s and 1950s. By Pew’s long-running measurement, in 1971 about 61% of Americans lived in middle-class households; the share of total income going to middle-class households peaked at 62% in 1970. That is the high water mark in this data series.

Lay the three periods on the same timeline: dumbbell ≈ 760 years, transition ≈ 180 years, bell curve < 80 years. The bell-curve society has existed less than 1/9 as long as the dumbbell. It is a historical exception, not the norm.

The bell curve was simultaneous medianisation across many dimensions

“Bell-curve society” is not just an income distribution. It is the result of industrialisation and mass culture medianising a society across multiple dimensions at once. Production medianised (industry concentrated huge amounts of goods in a middle quality / middle price band), aesthetics medianised (mass education and mass media compressed the aesthetic spectrum to a recognisable middle band), culture medianised (mass publishing, radio, and television turned mainstream culture into something most people shared), occupations medianised (large firms created enormous numbers of middle white-collar roles), educational paths medianised (standardised tests, college degrees, stable career ladders).

These five things are five faces of the same historical process, not unrelated coincidences.

The structure started collapsing in the 1970s. Mid-tier manufacturing roles, clerks, bank tellers, telephone operators, travel-agency staff, newsroom editors — these jobs disappeared from cities one at a time. By 2023, the share of Americans in middle-class households had dropped from 61% (1971) to 51%; the middle class’s share of total income had fallen from 62% to 42%. Of Americans born in 1940, about 90% out-earned their parents; of those born in 1980, only 50% did.

The mechanism that let the bell-curve society “reproduce middle class out of middle class” for a generation is failing. This matters: the collapse was not started by AI; it began in the 1970s and has now been running for half a century. What AI does is push a 50-year slow process to a new threshold.

AI is endogenous to the bell-curve society

Here is a key fact that gets overlooked: AI’s default commercial output lands in the middle band of mass expression and the mainstream English-speaking world.

The precision of this matters. The pre-training corpus of an LLM is far broader than its default output; everything from ancient Greek philosophy to top-tier mathematics papers is packed into the parameters. But most people’s interaction with AI does not happen deep in the parameters. What they touch is the default output that has been aligned by reinforcement learning from human feedback (RLHF).

The mechanism of that alignment is more precise than “graders’ aesthetic falls in the middle”. The people actually grading the model are not the average human. According to public information, OpenAI’s InstructGPT hired about 40 contractors via Upwork and Scale AI; in an anonymous survey, of 19 respondents 75% were under 35 and most came from the US or Southeast Asia. Anthropic’s 2022 RLHF paper used master-qualified US MTurk workers plus Upwork crowdworkers, with MTurk workers contributing the bulk of the comparison data and Upwork workers contributing the rest.

A more precise description: RLHF makes a model’s output converge toward “the region a stack of industrially recruitable, screened evaluators can stably identify as ‘better’ under researcher instructions and platform policy”. That region carries an obvious English-language, American, platform-shaped bias; it is neither the right tail of taste and judgment nor the global average. Anthropic’s 2023 GlobalOpinionQA study confirms this: aligned models default toward views closer to American and some European populations on subjective social questions.

Place this back into the history of the bell-curve society: AI is the bell curve’s endogenous accelerator, not an external shock.

Industry medianised production, mass education medianised aesthetics, mass media medianised culture, the large firm medianised occupations, and a stack of industrially scalable evaluators medianises AI output. This is the next link in the same historical chain.

This explains why AI accelerates homogenisation inside the bell curve, and it explains AI’s shared fate with the bell-curve structure: it can accelerate the takeover of everything inside the bell curve, but its default commercial output lacks the responsibility, judgment, and presence that come from real situations outside the bell curve.

---

Chapter 2 · Form Migrates Quietly

Chapter 1 set up the structure of the bell-curve society. Now back to AI’s form migration itself: how is this one different from the four before it?

There is a specific question to argue first: how fast does the public recognise that form has been absorbed by the tool, and how long does it take to repricing for that part of form to converge to the right value? In other words, how fast does the hand-producer of mid-band content lose value.

Public recognition speed depends on whether the tool produces a new form

Core claim: the speed at which the public recognises a new tool depends, at the core, not on its adoption rate but on whether the tool produces a visible new form or a distinctive change in the old form.

Historical samples:

• Print: produces a new form (printed page vs. manuscript look entirely different); recognition was immediate
• Photography: produces a new form (a photograph and an oil painting are different objects); recognition was immediate
• Industrialisation: distinctive change to old form (machine goods are too consistent); barely recognised at first
• Search: no new form (the searched content looks like things you already know); recognition lagged 15-20 years
• AI: at the deliverable level, no new form (AI-generated essays, code, slides, designs are visually indistinguishable from human deliverables); recognition is even slower

A qualifier first: this is at the deliverable level. At the product level, AI has produced new forms — chat-style interfaces, agent workflows, real-time generated video, multimodal collaboration spaces. These are recognisable, but what they recognise is AI as a tool, not AI as a mature tool of content production. The “form absorption” this essay discusses is at the deliverable level: when the public sits in front of an essay, a piece of code, a proposal, they cannot tell if AI or a subject made it.

For AI this time, public recognition has three possible paths:

Path 1: a new form appears. AI has a qualitative jump someday and produces a new form the public can directly identify (perhaps a new interaction, real-time generation, some entirely new shape). If this happens, recognition is immediate; the historical script repeats. But it is not guaranteed to happen.

Path 2: recognition by quantity. After AI content saturates, the public develops a statistical recognition through accumulated exposure (analogous to the dull “this is too consistent” recognition of industrial goods). It happens on a 15-20-year scale. But this recognition is not precise; what the public recognises is “AI style”, not “with substance vs. without substance”.

Path 3: never recognised. AI capability keeps improving, and its output converges with human work on every observable dimension. This cannot be ruled out. If it happens, the mass market never naturally identifies substance.

We are in the early stages of Path 2. The future could go any direction. This is real uncertainty, not the performative kind.

Even when the tool is universal, repricing still lags

Core claim: even when a tool is fully universal, people’s perception and pricing do not immediately fall back to marginal cost. The convergence is slow, usually on the order of decades.

Historical samples:

• Print: universal by the 1500s, repricing fully converged in the 17th century — lag of 50-100 years
• Photography: universal by the 1880s, repricing converged in the early to mid-20th century — lag of 30-50 years
• Industrialisation: universal in the early 20th century, hand-craft repriced in the 1960s-70s — lag of 50-70 years
• Search: universal in the 2010s, not yet fully converged (huge numbers of jobs are still hired on “knowing where to look”) — lag of 15+ years and counting; another 10-20 years for full convergence

The mechanisms of the lag:

• Stock-demand lag: existing customers of the old form keep paying out of habit and identity
• Hiring and salary system lag: HR and the talent market update slower than technology
• Education and evaluation system lag: what schools teach and assess can’t keep up
• Generational turnover: people trained on the old logic don’t reprice themselves; you have to wait for them to leave the market

The specific implication for AI: even after AI tools are fully universal, repricing will still take decades. If you stack Path 3 (never recognised) on top, repricing may never fully converge. This means the AI era could produce a transition period of 20-30 years, possibly never closing.

The two curves stacked: what does it mean

Stack the two together and the AI-era transition will be longer than any tool migration before it, and the endpoint is uncertain.

The actual current picture:

• Tools are not mature, or maturity is uneven: AI video six months ago and today are dramatically different; AI coding last year and this year too. Yet AI still hallucinates and errors frequently in places.
• Supply side: form-production cost has collapsed to the token level in the hands of those who can use AI (probably less than 1%) — this has already happened
• Recognition side: the public has no visible signal triggering recognition, possibly 15-20 years of slow quantitative recognition, possibly never
• Pricing side: even if recognition completes, repricing still lags by decades; full convergence may not happen within our professional careers

These stacked together produce a few claims:

• The public’s perceptual lag will continue; tool maturity is uneven; recognition starts asynchronously across fields
• Form production is collapsing to the token level, but the market is still pricing people on the old logic
• The old logic still keeps some people alive — on market lag, not on personal ability
• During the lag, who gets repriced first by their clients/employers vs. who gets repriced later is mostly luck. Two people of equal ability can land in completely different places.
• Most people will sense an “I-can’t-quite-name-it wrongness”: prices slowly compress, but without any obvious crisis signal

This gives knowledge workers and firms a genuinely existing transition period.

The monthly salary still arrives, clients still pay, hiring still runs on the old logic. This time is real. It is the result of the market updating slower than the tool, not a hallucination. It opens some space: re-position, accumulate substance, adjust direction.

But be clear about one thing: this time is a transition, not a new equilibrium. The supply side has already collapsed; the market just hasn’t filled in recognition and pricing yet. How long this lasts depends on your field, your client type, market inertia; there is no single answer, but it has an end.

Using this time well, and mistaking this time for a steady state, are two different things.

One open question

Here is the actual landing point of this chapter:

This is the first time in history that, after a tool migration, the public may not be able to recognise it directly.

In each previous tool migration, the public eventually recognised it: through a new form, through accumulated exposure, through quality of results. AI may be the first tool that has none of these channels, or that bypasses them all. Its output and human work look the same on the surface, and the capability keeps growing fast — the gap may keep narrowing rather than widening.

This means the public may not be able to tell whether the form of a thing was supplied by AI or by a subject.

So can the public still reliably identify substance?

---

Chapter 3 · Substance Is Everywhere; Few Identify It

The previous question: can the public still reliably identify substance? The answer is not a clean “yes” or “no”. Start with a fact that looks contradictory.

The mass market pays every day for an enormous number of products and services — Meta’s apps, Uber, MrBeast, Apple, Netflix — and most of these have real substance behind them. But the public never directly identifies the substance; they consume something else.

This looks contradictory; in fact, it reveals a picture more precise than “recognise vs. don’t recognise”: substance is everywhere, but the people who can identify it directly are very few. To see this clearly, look at what is actually happening in deliverables that get paid for today.

Currently-paid deliverables × substance type × recognition level

Eight representative deliverables, with notes on whether substance exists, what type it is, whether the buyer can identify it (i.e. know what they are paying for), and what they are actually paying for. Recognition is marked in three tiers: directly identifiable, indirectly identifiable (outcomes, long-term use, word of mouth), essentially unidentifiable.

Two rows in this table deserve a pause: “mass apps” and “internal enterprise deliverables”. One is the consumer market, one is the B-side market. In both, the central fact is the same: there is real substance, but the buyer is not paying mainly to identify it.

What the table tells us

Claim 1: substance is everywhere

A substantial proportion of currently-paid deliverables carry real substance. The network-type ones (Meta apps, Uber), the node-type ones (MrBeast, Apple), the embedded ones (Visa, SWIFT) all contain something real.

It is not “post-AI, the world is all hollow shells”. There are plenty of hollow shells, but plenty of products carry real substance behind them.

Claim 2: the public rarely identifies substance directly

This may break a few priors, but it is the actual situation: the public cannot articulate the substance; they can only feel its projection.

The public consumes substance, but what they feel is a low-dimensional projection of it, not its full structure.

• A user using Uber feels “I open the app and a car comes” — this is the projection of Uber’s two-sided market plus its algorithm
• A user on Meta’s apps feels “all my friends are here” — the projection of the social graph
• A user buying an iPhone feels “it just works” — the projection of Apple’s product judgment
• A viewer watching MrBeast feels “this video is a blast” — the projection of MrBeast’s judgment plus industrial execution
• An ordinary consumer choosing Dunkin’ or Starbucks is choosing “cheap, drinkable, reachable” — but those three things are projections of supply-chain control, real-estate and store-build capability, and people-management capability (substance)

In every case, the public consumes real substance — they just see the projection rather than the thing itself.

This is more precise than “the public doesn’t recognise substance”. It admits that the public really does pay for substance (through the projection), without saying the public is “fooled” or “consuming hollow shells”.

There is a precise economic correspondence to this. Nelson (1970) and Darby & Karni (1973) divided goods by the consumer’s ability to identify quality into three classes: search goods (visible before purchase), experience goods (knowable after use), and credence goods (still hard to judge after use). Substance is fundamentally a credence good: when consuming an essay, a product, or a service, the public cannot reliably distinguish “real subject judgment” from “AI default output”. Akerlof’s 1970 lemons model (Nobel 2001) proves it: in a market where quality cannot be identified, quality itself cannot be priced.

Claim 3: enterprise-market recognition ability is highly uneven

The B-side is in theory better at identification than the public, but in practice splits in two:

• The minority of high-taste decision-makers (clients who actually understand the business, senior partners, technical experts): can identify directly
• Most mid-tier decision-makers: identify mainly form — whether the deck looks good, whether the report is complete, whether the deliverable is on time; some have no standard at all and rely on intuition and personal preference

The mid-tier decision-maker’s only old tool was form signal. AI breaks that tool, and they have nothing to replace it.

People at the top of the taste curve do not just see the projection; to a degree they can also infer back to whether substance exists. They look at a proposal and feel “this is not AI-flavoured speculation, this is someone who has worked through similar problems before”; they look at a piece of code and identify “there is judgment in this architecture, it isn’t just stacked”; they hear a song and hear “there is taste in this arrangement, it isn’t templated”.

The essence of taste is the ability to “infer substance from projection”.

The public lacks this ability. They stop at the projection layer.

Does AI affect this chain?

It does, deeply.

In the past, the only path to producing the projection was almost always through substance. A cheap coffee with a good projection had to have a real supply chain behind it. A product that “just works” had to have real craft. A proposal that lands had to have real experience. Projection is a low-dimensional rendering, but it reliably stood in for substance.

AI changes exactly this. It makes the cost of “fabricating the projection” close to zero, and the output quality in every domain corresponds to the upper-mid of the bell curve, beyond what most people can distinguish. AI can produce an article that “looks like there’s judgment in it” without any judgment. It can produce a product demo that “looks usable” without anyone actually understanding the user. It can produce an analyst’s report that “looks professional” without anyone having worked on a similar problem before.

These are not projections: there is no substance behind them. They are mirages: they look like projections, but the tool produced them directly.

A concrete example: the interview

An example.

You interview a candidate. The CV is clean, the answers are crisp, the cases are apt, they react quickly to your questions, they articulate “an interesting view” of the industry. You form a judgment: this person has the chops.

The chain of that judgment is: chops (substance) → CV and answers (projection) → your hiring decision.

This chain used to work because producing this projection itself required corresponding substance: a clean CV needed real prior work, crisp answers needed real understanding, apt examples needed real experience. The projection is a low-dimensional rendering, but it reliably stood in for substance.

AI rewrites every step of this chain. AI helps draft the CV, prep the answers, run mock interviews; in remote interviews it can even feed real-time prompts off-screen. A candidate without the substance can, through tools, produce almost the same projection as a candidate with it.

The “clean CV”, the “crisp answers”, the “apt cases”, the “interesting view” you see may be the projection of substance — or the mirage of AI. You cannot tell.

There is a deeper layer. You will be inclined to believe what you see is the projection. Because if you accept that “this could be a mirage”, how many people in your last few years of hires came in like that? It is a question you don’t want to open.

This is not a new phenomenon unique to the AI era. Spence’s 1973 signalling theory points to the structural fact: when quality cannot be observed directly, the market relies on signals to identify it; but the only mechanism that lets a signal sustain the distinction is mimicry cost — low-quality producers must pay a high cost to mimic the signal. AI compresses this mimicry cost to nearly zero, and the signal system fails structurally.

A flood of artefacts

More cases like this are happening or about to happen.

This means that for a bell-curve society (public taste in the middle, AI floor at upper-mid bell), the upper-mid AI floor produces a flood of artefacts on top of the substance projection.

“Artefact” here is borrowed from medical imaging and signal processing, where it means something that looks like real signal but isn’t. Projection is real signal (low-dimensional). An artefact is fake signal that looks like a projection.

The bell-curve society faces AI from a doubly disadvantaged position: public recognition is in the middle to begin with, and the AI mirage quality lands exactly in the band the public cannot distinguish. The public’s recognition system is being occupied by artefacts; this is a structural failure, not just a lag.

The consequence runs both ways:

• People without substance use AI to fabricate projections, and in the mass market they look the same as people with substance
• People with real substance produce real projections, and in the mass market they look the same as the tool’s artefacts

The two structurally merge at the level of public perception.

So where does the value of substance get clearly identified?

---

Chapter 4 · The New Filter: From Form Quality to Substance Origin

The only place substance gets clearly identified is in the high-taste market — the slice of the audience that can infer substance from projection.

This is something history has done before, not a new mechanism. When a signal is widely mimicked and loses its discriminating power, the market that can discriminate spontaneously upgrades to harder-to-mimic signal systems. This is a regularity repeatedly observed in consumer behaviour research. Feltovich et al. (2002) on counter-signalling and Bellezza (2023) on distance-based alternative signals both make the same point: when the middle-tier participants can use the mainstream signal too, the top tier shifts to “signals that don’t look like signals”, or to markers that require more inside knowledge to identify.

Quiet luxury is one concrete shape of this: once the logo is widely mimicked, taste-led consumers move to “no-logo high quality only insiders recognise”. The same mechanism unfolds in the AI era: once “form passes the bar” is mass-mimicked by AI, the high-taste end of the market shifts to new, harder-to-mimic signals — subject identity (who made it), real track record (past judgments and accountability), real recommendation among insiders (transmitted trust).

This is a partial but available historical map. We can only do guesswork from it for now, sketching one possible outcome.

After print, authorship became a new organising principle

Before print, most texts were anonymous, attributed to authoritative tradition (Aristotle says, Scripture says), or passed down as collective copies. After print, “who wrote it” became the central organising principle of the market: readers filtered by author, the author became a kind of brand. Foucault’s well-known essay “What Is an Author?” is about exactly this — author as a modern category came into being only after print.

The AI era is replaying that process. Once passing-the-bar form is no longer a scarce signal, the filtering mechanism migrates again toward “who made it”. This is a structural shift with historical precedent.

But the filter only works on part of the market

A qualifier here. This is a trend in motion, not a finished state. The terminal state takes shape on a roughly 10-year horizon.

The terminal picture: in every domain, subject-identity filtering only works on the high-taste audience. The mass audience consumes anonymous content; what changes is that the source of that content shifts from a large pool of human creators to AI.

Today is not the terminal state. Vestiges of subject-identity filtering still exist in the mass market: fans following stars, readers following authors, users following brands. This layer is rapidly thinning: an internet-celebrity life-cycle has compressed from years to months; brand stickiness is dropping; mass memory of and devotion to “this specific subject” is getting thinner.

Another feature of the terminal state is a split into two kinds of “subject identity”:

• Capital-grade IP: mass IP built by huge capital investment, algorithmic distribution, and continuous industrial production (Disney, top game studios, top-tier platform celebrities). This kind of subject identity persists, but it is the product of industrialisation, not a path most founders or individuals can reach
• Author-grade subject identity: identity built up by real judgment and track record, recognisable to the high-taste end of the audience. This is the kind this essay discusses

In the terminal state, the middle layer — “an ordinary subject going to mass scale on personal charisma alone” — structurally disappears, because AI makes mass-content supply infinite and ordinary individual charisma can no longer be reliably identified inside the supply flood.

The print era was actually the same. Authorship as filter worked on elite readers: intellectuals filtered classics by author; the academy traced arguments by author. But for the mass reader, the author mechanism has always been limited. Most pamphlets, popular religious texts, and cheap adventure novels of the 16th to 19th centuries were anonymous. The high-taste end filtered by authorship; the mass market consumed anonymous default output — except this time, “anonymous” is “AI”.

For the high-taste audience, the new filter has three layers

The new filter is not single. It has three layers.

Dominant layer: subject identity (authorship). In a supply-flooded environment, readers, users, and clients no longer have time or capacity to filter from the content itself. They filter by “who made it”. Any output by a subject with a traceable track record automatically enters the reader’s attention; any output by a subject without one, however good the content, does not. The essence of this layer: what readers actually consume is “I trust this subject’s judgment”, and the content is just the carrier of that judgment.

Second layer: relational signal. Readers filter via “recommended by someone I trust”. In a supply-infinite world, both algorithmic recommendation and ad placement are diluted by AI content; real personal recommendation becomes more important than before. But this layer is built on top of subject identity: content without an identifiable subject can’t be carried in the trust network either.

Third layer: scarcity signal. Lavish packaging, limited editions, paywalls, invite-only — these strategies have worked historically too. But they are niche tactics for a few players, not the dominant organising principle.

The weights of these three layers have shifted from before. Form quality used to dominate; subject identity was the second layer. In the future subject identity dominates, and form quality drops to a threshold (below it you’re out, above it the differences stop mattering).

What this means for two kinds of readers

What follows is the 10-year terminal state. The old system is still running today, but the pressure of the terminal state has begun to show.

For knowledge workers: the old way of running a career was “get form to passing” — write a passable report, deliver a passable proposal, ship a passable design draft. That path has collapsed.

The new path: “build a subject identity recognisable to the high-taste end of your field”. Let that part of the market filter by “who made it” rather than “how well it’s done”.

What that looks like in practice: public output, leaving traceable judgment records, sustained presence in a single domain for long enough. These were “extra credit” before — you produced public content, attended industry conferences, and made occasional public statements on the side of your day job. Now they have become the only thing that counts; extra credit has been upgraded to required.

A knowledge worker who does not build a subject identity is invisible under the new filter. The problem is not whether they are good; the problem is that they are not on the filter’s working surface.

For start-ups: the old way of running a product was “find a gap in form/function and fill it” — make a tool, an app, a SaaS. This path is shrinking fast, because AI fills any function gap quickly.

The new path: “make the product carry subject identity”. A recognisable subject (or a small group of subjects) stands behind the product; users pay for “the judgment of this subject”, not just for the function.

This explains why so many product successes in the AI era have a strong personal IP as the front. It is structural necessity, not a marketing trick: a product without subject identity has no reason to be chosen in a supply-flooded environment.

But one thing must be emphasised right away: the mass market and the high-taste market pay for subject identity in different ways.

The mass market pays for a consumable projection of identity. Stars, internet celebrities, athletes, game studios, luxury brands, Apple, Nike, Taylor Swift, Hayao Miyazaki, miHoYo — these are concrete examples of mass-market identity-monetisation. But what the mass consumes is the projection of identity (halo, belonging, emotional connection, brand signal), not the subject’s judgment itself. Monetising identity at mass scale generally requires going through capitalisation, entertainmentisation, branding, or fan-relationship machinery, packaging the judgment into something consumable.

The high-taste market can pay directly for the judgment itself. They can identify “the depth of this subject’s judgment” and are willing to pay a premium for it directly, not for its derivative packaging.

For a start-up this means two different paths: a mass-market path requires the “projection-monetisation” route (capital + industrial production + fan-relationship management); a high-taste-market path can monetise directly on judgment. The middle path — “mass product, hoping the mass market identifies my judgment itself and pays a premium” — is structurally hard to walk, because mass-market recognition precision does not extend that far.

Choosing the field matters more than choosing the work

The framework yields one sharp recommendation: choosing the field matters more than choosing the work.

The economic value of a subject identity depends on two things:

• How good the subject is (how deep their judgment, taste, track record)
• How large the high-taste audience that can identify them is, in their field

The first you accumulate yourself; the second is set by the field. The product of the two is the subject’s monetisation space.

Two common wrong choices follow:

One: doing extremely well in a field with an extremely small high-taste audience. The subject may be admired in the professional circle, but the monetisation space is bounded because the people who can identify them are few to begin with.

Two: passing-the-bar inside a mass-market field. This used to be the safe choice — large audience, passing was enough to eat. In the AI era this path collapses: the economic value of “passing” goes to zero, the mass audience does not identify the difference, and subject identity does not monetise here.

The ideal position: a field with a high-taste audience large enough to support you, doing work that part of the audience can identify. Such fields include some popular categories with hard-core fan communities (web-fiction with serious readers, games with a connoisseur community, technology or finance content with deep readers) and some professional areas (B2B niches with enough paying capacity).

Not every field has a large enough high-taste audience. But when picking direction, the audience structure of the field matters more than your own interest. Hard work in the wrong field can’t beat passing-the-bar work in the right one.

---

Chapter 5 · The Two Carriers of Substance

The two most structural moats in the AI era: authorship (node-type) and embeddedness (network-type). Everything else is depreciating.

The structure is fully symmetric across the firm scale and the individual scale. One theory covers both.

Why substance has only these two carriers

AI can generate or replicate everything else: form, craft, information sorting, standardised execution, generic judgment, generic functionality. Those things used to be expensive because producing them took time and skill. AI drives the production cost to zero, and in the new boundary-drawing they all fall under form.

What AI can’t replicate, looked at carefully, is only two kinds. These two are the carriers in which substance can reside:

Node-type (authorship): a recognisable identity built by a specific subject through long-term presence, accountability for outcomes, and a traceable record of judgment. Substance resides in a specific subject.

AI has no specific subject, no history of bearing consequences, no judgment trajectory calibrated against reality. It can imitate “what this subject’s judgment style would look like”, but it doesn’t have that subject: no validated stretch of time, no traceable list of judgments, no real consequences borne. This kind of substance cannot be replicated by tools.

Network-type (embeddedness): a network built up by a set of real relationships running together over time. Substance resides in the joint operation of a set of relationships. “Relationship” here is broad: it includes real interpersonal/client relationships as well as presence in specific physical infrastructure (logistics networks, fabs, data centres, energy grids) and institutional embeddedness (regulatory licences, compliance qualifications, government procurement status, patent portfolios). What these have in common: they come either from real relationships built up by long-term operation, or from positions inside specific physical or institutional structures, and AI cannot generate them out of nothing.

AI can simulate connections, but it cannot synthesise real shared experience, mutual calibration, and high-bandwidth trust. It can produce a list that looks like “a relationship network”, but it doesn’t have that set of relationships: no shared history of getting through hard things, no resumed collaboration after a failure, no “you don’t have to explain it” understanding. This kind of substance, again, no tool can replicate.

Everything else — surface brand, channel position, scale, generic data, generic tech stack, generic functionality — either is form, or sits on top of form, and is structurally depreciating in the AI era. They looked like “substance” in the past only because the old tools weren’t strong enough; this part of form was being sold as substance. Once AI arrives, the disguise breaks.

Only node-type and network-type assets actually carry substance, and AI cannot touch them.

Side-by-side at the firm and individual scales

Putting the two moats side by side at both scales:

The two moats operate by completely different mechanisms, but share one feature: AI cannot touch either of them.

Node-type is centralised: value is concentrated in a specific subject. Replace the subject, and the value disappears.

Network-type is distributed: value is distributed across a set of relationships. Any single node can leave, and the network is still there.

The two moats face different fates in the AI era

Network-type (embeddedness) does not automatically appreciate in the AI era from filter migration. AI mostly cannot touch Amazon’s logistics, Meta’s social graph, Visa’s payment network. These things are not weakened by stronger AI, because they don’t sit on the surface where form is taken over by AI. Their appreciation comes from network expansion itself, or from AI enhancing operational efficiency (better recommendations, more stable risk control, more efficient scheduling), not from the dividend of the filter shifting toward subject identity.

Node-type (authorship) appreciates relatively in the AI era. Because the market’s filter migrates from “form quality” to “subject identity”, and AI collapses the form-quality dimension, the relative value of authorship rises. The same authorship that was worth X before AI may be worth 3X or 5X after. What changed is its relative importance in the market; the strength of the authorship itself didn’t move.

Which means: both moats have value, but node-type is the type that genuinely benefits relatively in the AI era; network-type is the floor.

Time is necessary, not sufficient

Both moats require time. But time is necessary, not sufficient.

• Authorship needs time: traceable judgment requires sustained public output, track record requires long-term accountability, presence requires staying around
• Embeddedness needs time: a set of real relationships running together requires long-term joint work, mutual calibration requires multiple rounds of real contact

But time alone isn’t enough. People who have only time discover their time doesn’t monetise: they may have long tenure, but they haven’t built a recognisable subject identity, and they haven’t built a deep relationship network. Their time becomes a sunk cost.

You don’t need both moats at once

A lot of people instinctively assume “the best companies have both authorship and embeddedness”. But in reality, a firm with only one of the two moats can be excellent.

A24 is almost pure authorship — it’s a distribution brand, with no significant embedded assets. Visa is almost pure embeddedness — nobody says Visa stands for some kind of “judgment” or “taste”. Both are top of their fields.

The same holds at the individual level:

• Authorship only: independent thinkers, independent writers, independent consultants. They may work alone behind closed doors, but they continuously produce recognisable judgment
• Embeddedness only: serial founders, senior connectors, industry veterans. They may never publish anything publicly, but they’re irreplaceable inside a set of deep relationships

Both paths work. The key is to be conscious of which one you are running, and act accordingly.

The worst case is trying to run both without doing either well. That’s the reality of most middle-tier knowledge workers.

One last qualifier

Both moats, in the 10-year terminal state, share one prerequisite: they only matter to the part of the market that can recognise them.

Authorship needs an audience with taste before it can be recognised as valuable. Embeddedness needs buyers with judgment before it can be recognised as irreplaceable.

Which means both moats are locked into the same audience structure as “the high-taste end”: they are both assets of the high-taste end of the market. Outside that, neither moat produces economic value.

So while you build either moat, you also have to find the part of the market that can recognise it. If you can’t, the moat is a sunk cost.

---

Chapter 6 · What This Means for Firms

At the firm level, the two moats yield a concrete asset taxonomy: which of a large firm’s assets are authorship, which are embeddedness, which are pure form, and what this means for different kinds of firms.

Three classes of firm assets

AI forces the previously fuzzy concept of “corporate moat” to be classified. The simple test: could another firm, without owning this firm’s specific history, generate the same asset out of nothing? If yes, it’s form; if no, it’s time (which further splits into authorship and embeddedness).

Walking through the traditional moats item by item:

One: pure-form assets (accelerating depreciation)

These were treated as moats but AI forces them to expose themselves as form:

• Generic tech stack: architecture, libraries, tooling that took hundreds of people years to build. AI can replicate the same functionality scale in weeks
• Platform-dependent channels: search rankings, App Store position, social-algorithm distribution, paid acquisition accounts. They depend on the current state of platform algorithms; one platform tweak resets them
• Dead IP: expired patents, dormant trademarks, retired characters — replicable or routable by AI at any time
• Generic data: data that mostly repeats public-corpus distributions; AI can derive something similar without spending the time

These assets will systemically depreciate over the next 5-10 years. Many large firms still treat them as moats in internal valuation and continue to amortise them on their books, but their economic value is already falling.

Two: embeddedness assets (holding value)

Relationships embedded between the firm and users’ day-to-day operation, which AI can’t touch:

• Logistics networks: Amazon, SF Express, UPS last-mile delivery and dispatch
• Social graphs: Meta’s, LinkedIn’s real interpersonal networks
• Core infrastructure: Visa’s payment rails, SWIFT’s interbank clearing, Google Maps’ real-time geographic data
• Embedded proprietary data: 10 years of hospital case records, 20 years of insurer claims, a logistics firm’s anomaly-response logs — produced from real operation, not present in public corpus

These assets don’t hold value on their own. They are by-products of the property of “being present”: presence yields data, no presence means no data.

In the AI era they don’t automatically appreciate from filter migration, but they may appreciate from AI enhancing operational efficiency. Better recommendations, more stable risk control, more efficient scheduling — all step up the operational efficiency of network-type assets.

Three: authorship assets (relative appreciation)

Recognisable subject judgment, in firm form:

• Apple: product trade-offs, aesthetic style, judgment about user relationships (years after Jobs’s death, the judgment framework he left is still operating)
• A24: taste in selection, distribution style, support for auteur cinema
• Berkshire Hathaway: Buffett and Munger’s investment-judgment framework
• Patagonia: environmental stance, product philosophy, supply-chain trade-offs
• Aesop: aesthetic style, taste in store location, consistent judgment about product experience

These firms share one feature: every decision they make can be traced back to a recognisable judgment framework. Touch any one point, and you can identify “this is from the same firm”.

They appreciate relatively in the AI era. Because the filter migrates toward “subject identity”, their relative importance rises systemically.

Three kinds of firm, three prescriptions

By asset mix, firms split roughly into three classes, each with a different fate and response in the AI era.

Class one: large firms with embeddedness but diluted authorship

This is the situation of most legacy giants. Amazon, Google, Meta, Microsoft, the big banks, the big telcos, the big retail chains — they have strong embeddedness, but their authorship has been systematically diluted after founders exited and professional management took over.

Their biggest structural weakness in the AI era is authorship. Embeddedness protects them from being replaced quickly, but they cannot capture the dividend of filter migration, because they have no recognisable judgment framework. Users use their products because there’s no choice (high switching cost, strong network effects), not because they like it.

The risk is not today; it’s on the 10-20 year scale. Once embeddedness gets routed around (new tech, new regulation, new user habits), they have no authorship to fall back on. A few firms patch this through founders or strong subjects refusing to retire, spinning out independent sub-brands with authorship, or acquiring and preserving small firms with authorship — but most large firms won’t. They will keep coasting on embeddedness, while losing relative position to authorship-driven newcomers.

Class two: startups (structural advantage in authorship)

Before the prescription, one sharp thing: most of what the past 20 years of VC models treated as “competitive advantage” does not constitute a moat in the AI era — and it’s not slow attrition, it’s accelerated collapse.

Product UX, technical architecture, craft mix (the React + Tailwind + Supabase + Vercel kind of stack), product ideas and features, public or basic data, brand packaging — all the things that used to take millions of dollars and years to do at a passing level can now be made in the same form by anyone in weeks with AI.

Sharper still: mimicry. A new firm with nothing can use AI to produce, in two weeks, a product demo, website, case studies, and customer testimonials that look just like yours. They don’t have to actually build it; they only have to fake the projection, and median decision-makers (investors, clients, partners) can’t tell real replication and fakery apart.

Which means the core narrative of startups in the past — “we built X to get Y, so we’re worth this valuation” — has had its X→Y causal link broken by AI. Others can fake Y without doing X.

The core VC narrative of the past 20 years — “build a better product experience plus tech plus data, then scale” — has every layer flattened by AI. The marginal cost of “better” collapses to near zero, and so does the marginal cost of “looks better”.

These assets won’t disappear overnight. The transition period, as discussed earlier, is real. Salaries are still being paid, contracts still signed, products still in use. But the relative value of those assets is falling fast. A startup still raising and operating on the logic that “our tech stack / product experience / data / craft mix is the core moat” is raising on a depreciating asset.

What’s left is what the previous chapter named: recognisable subject judgment (authorship) and relationships embedded in real operation (embeddedness). Everything else is quicksand.

So a startup’s strategic move is to systematically out-run large firms on the authorship dimension. Concretely:

• Make every product decision let users recognise the subject’s judgment. Don’t make the “generic, what everyone does” product; make the “we believe it should be this way” product
• The founder is on the front line, not hidden behind the product. Public expression, accountable judgment, traceable decisions — this is part of the product itself, not just marketing
• Actively choose the high-taste end, not the mass market. The mass market doesn’t pay a difference premium; subject identity doesn’t monetise there
• Accept small scale, high price, deep relationships as the economic model. Drop the old script of “scale to millions of users fast”

The biggest trap: startups most easily err by mimicking the large-firm path: chasing scale, generality, standardisation. Those were correct in the VC model of the last 20 years, but that model was built on a “form is scarce” economic base, and the base has changed. “Scalable” in VC logic is a counter-signal in the AI era, meaning the product has no inseparable subject judgment in it.

Class three: firms with neither moat

This is the biggest and most overlooked class. A great many mid-size SaaS firms, traditional consumer brands, professional service firms, and sellers on platforms — they used to rely on form scarcity (passing-grade product, passing-grade service) plus organisational embeddedness (stable team and process), and both disappear in the AI era simultaneously.

These firms won’t vanish immediately, but they’ll be systemically hollowed out over 5-10 years. Because:

• Embeddedness isn’t enough to protect them: their “embedding” is really switching-cost laziness, not real network effects or infrastructure
• Authorship is absent: no recognisable judgment framework, the product is a stack of “industry-standard practices”
• Form is no longer scarce: AI drives the production cost of homogeneous competitors to zero

Two paths out:

• Add authorship: founder returns, build a recognisable judgment framework, redo the trade-offs. But this requires management willing to give up the inertia of “stable operation”, which most can’t
• Add embeddedness: dig deep into users’ day-to-day operation, form non-detachable relationships. But this requires giving up the “spread the net wide” scale fantasy, focusing on a few truly deep client relationships

Do neither, and you depreciate systemically over 5-10 years.

A few old assets that aren’t moats anymore

A few counterintuitive judgments fall out of the classification, worth saying directly.

One: brand is not an automatic time asset. Brand was treated as a typical time asset — the longer it accumulated, the more valuable. But in the AI era, brand has to be continuously sustained by real action. Maintaining a brand with AI-generated content, diluting the product to push volume, papering over product problems with marketing — all of this degrades the brand into a form asset (a name, a logo, a template) within a few years. The market will reprice the degraded brand to the cost of its formal parts.

Two: scale is a multiplier, not an asset. Scale amplifies whatever real asset is underneath. Amplifying embeddedness (logistics serving a billion users) — scale has value. Amplifying form (a homogeneous app serving a billion users) — scale depreciates with the form. Scale itself doesn’t preserve value. Many large firms’ scale was built on serving the mass market. The per-user value of the mass market is being structurally compressed in the AI era; the multiplier of scale acts on a shrinking numerator.

(Tech stack, craft mix, generic data — discussed in the startup section — apply equally to large firms. They are the fastest-depreciating asset class in the AI era; not expanded again here.)

One concrete action

For a firm leader (CEO, business head, board member), the takeaway action of this chapter is:

Audit your firm’s asset list and reclassify into the three classes.

Write down everything currently treated as a “moat” or “core asset”. For each, ask the test: could another firm generate this without owning our specific history? Sort the answers into three piles: pure form, embeddedness, authorship. Look at the relative sizes.

Most firms doing this audit will find: the pure-form pile is far bigger than they thought, and the authorship pile is far smaller. The diagnosis isn’t wrong — it’s that the old accounting and strategic vocabulary mixed the three classes together, hiding the picture.

The choice afterwards depends on which of the three classes the firm belongs to.

---

Chapter 7 · What This Means for Individuals: Which Class You Are In

The previous six chapters gave the mechanism and macro judgment. Back to a concrete question: for a specific knowledge worker, what does all this mean?

To answer it, sort out one thing first: in the AI era, an individual’s situation can be classified along one underlying dimension that has nothing to do with profession, industry, or age. The dimension is: is your judgment something other than what AI does?

AI splits people into two classes: amplified vs replaced

AI’s effect splits sharply across people. For some it’s an amplifier, magnifying what they can produce. For others it’s a substitute, squeezing them out of the market.

What decides which side you fall on? Where your bottleneck is.

• Bottleneck is insufficient experience and knowledge (you want to do something but can’t, because you haven’t seen, don’t know, haven’t learned): AI is an amplifier. AI fills in what you can’t reach, and you use judgment and taste to combine them into a product beyond your own scale. This person has upper-layer judgment but lacks lower-layer information and tools; AI gives the lower layer, the upper layer monetises through amplification
• Bottleneck is not experience and knowledge but lack of upper-layer judgment: AI is a substitute. Their old moat was “I know more, I’ve seen more”, and AI flattens that. Most of what they know is already in the public corpus, and AI can recite all they know — and integrate it faster

The key isn’t “more or less knowledge”. It’s whether there is something further upstream of knowledge: judgment, taste, the ability to question the problem itself, the ability to abstract a specific experience into a structure. These upstream abilities are the real substance.

This is the same thing as the cognitive decoupling ability in Issue 01. Cognitive decoupling is the ability to lift a problem out of its specific context and handle it with abstract structure. For people who can do this, experience and knowledge are raw materials, and substance sits above the materials. For people who can’t, experience and knowledge are everything they have; and AI now also has the materials.

AI amplifies “people whose substance exceeds their form”, not everyone. For people whose substance equals (or nearly equals) their form, AI doesn’t amplify — AI replaces.

The economic correspondence is rigorous

This isn’t intuition. There’s a precise framework in economics.

The standard tool for “is a technology a substitute or a complement” is the elasticity of substitution (σ). Whether AI is a substitute for or complement to a person is determined by this parameter:

• σ > 1: AI substitutes for this person’s ability easily; AI deployment lowers their value
• σ < 1: AI is a complement to this person’s ability; AI deployment raises their value

“Bottleneck is experience and knowledge” maps to σ < 1: the person’s upper-layer ability (judgment, taste) and AI (supplementing knowledge, integrating information) are on different dimensions; AI amplifies the person. “Only experience and knowledge” maps to σ > 1: person and AI on the same dimension; AI substitutes directly.

The latest research repeatedly verifies this. A study based on 12 million US job postings from 2018-2023 shows that AI-related jobs demand complementary skills like resilience, agility, and analytical thinking nearly twice as often as non-AI jobs; these skills carry wage premia; the complement effect in the data is up to 50% larger than the substitution effect.

In other words: AI makes upper-layer ability more valuable and lower-layer ability less valuable.

Closing the loop: experience and knowledge go from moat to sunk cost

This mechanism yields a cold corollary: knowledge workers who currently have “only experience and knowledge” will be displaced. The displacers are people with judgment, not new entrants per se.

Two paths bring those displacers in:

• Same-generation peers with decoupling ability. They have judgment, plus all the embeddedness this worker has: clients, relationships, industry familiarity. They can use AI to amplify like a new entrant, and they have embeddedness new entrants don’t
• New entrants directly identified by high-taste employers. They have no accumulation, but they have judgment, and they bypass embeddedness barriers via direct identification by employers (without going through resume filtering). Before AI this was nearly impossible, because employers had no time or tools to identify judgment directly and had to lean on resumes as second-order signals. After AI, with resumes polluted by AI and the cost of public output of judgment crashing, employers gain the ability to identify judgment directly, jumping over the resume

Both paths route around the same thing: experience and knowledge themselves.

When the filter migrates from resume to judgment, experience and knowledge go from moat to sunk cost.

The painful part: most people currently in white-collar knowledge jobs will find they are in the “only experience and knowledge” class. This is not because they didn’t try. Past education, career paths, and corporate organisations were systematically training this class: middle management at large firms, skilled workers in professional services, senior analysts — these positions’ value proposition was precisely “accumulating experience and knowledge”, not judgment and taste. Standardised education rewards “mastering a lot of knowledge”; corporate KPI systems reward “stably and reliably doing the job”; neither develops independent judgment.

These people are the bell curve trained by the era. Their pain will be especially sharp because they will realise they did everything they were taught to do, and ended up at a dead end. They walked the manual step by step, every step met expectations, and at the end the manual itself was voided.

Self-audit: which class are you in

Here, the question that really matters is: which class are you in?

This almost cannot be identified from the inside. Identifying “do I have upper-layer judgment” itself requires upper-layer judgment. A person without decoupling ability typically doesn’t realise what they’re missing; the world they see is “experience and knowledge matter most”, because that’s what they themselves run on. The person who most needs this judgment is precisely the person least able to make it.

Three positions to compare:

• Judgment is real: accept the reality of slow monetisation and long-term sustained presence; don’t pretend you can grow fast. This person is slow but on a real path
• Judgment is genuinely absent: accept that you’re on the luck track, don’t pretend to be building a long-term moat. This person is fragile but at least their decisions point in the right direction
• Thinks they have judgment but doesn’t: makes long-term decisions on wrong expectations, refuses luck-track form-grinding opportunities, and pours large amounts of time into “substance work” no one identifies. Misses on both sides

The third is the most dangerous position.

The only externally calibratable check is: whether your substance has been recognised and responded to by people with taste. The standard is sustained recognition by people whose judgment you yourself respect, in your field. Praise from acquaintances, recognition from coworkers, one or two clients buying — none of those count.

Without that response, with high probability you’re overestimating yourself.

This is cold. But better to know early than burn 10 years in the wrong position.

Three classes of reader, three different situations

After the underlying mechanism, the three classes of reader face very different situations.

One: judgment + accumulation (this essay’s true target reader)

You are the class AI amplifies. AI lets you do an order of magnitude more than before. What you should do is run authorship and relational capital up to the point where the market can identify them, not figure out how to “defend against AI”. Specifics in the next chapter.

Two: accumulation + no judgment (the middle tier)

You are in the most complex position. On one hand, embeddedness still protects you: client relationships, industry familiarity, internal company position — AI can’t touch these in the short term. On the other hand, structural depreciation has begun: same-generation peers with judgment, plus new entrants identified by high-taste employers — both paths are closing in on you.

The middle tier won’t vanish overnight. For a long time, the middle tier will continue to exist, but its mode of existence shifts from structure to crevice: a niche AI hasn’t covered yet, a client still paying on old logic, a position temporarily exempt for compliance or inertia, an instance of personal charisma or luck landing right. These are crevices, not structure. You can survive in crevices; some people last a long time. But they are unpredictable, untransmissible, non-reproducible.

The most dangerous thing is not realising you’re in structural fragility. Salary still arrives every month, life goes on, no obvious crisis signal. Until one day the window closes (layoff, industry shrinkage, client loss), and you suddenly find you have nothing.

If, on honest examination, you find yourself in the middle, this paragraph offers one piece of advice: treat the time inside the crevice as a window, not as steady state. The middle tier’s income gives you time and resources to build authorship and relational capital (people not in the middle don’t have this luxury); the middle tier’s position gives you contact with real problems (if it’s a real training ground, not a biological API). Those who can see this honestly and act on it can use the window to prepare. Those who can’t will be caught flat-footed when the window closes.

Three: judgment + no accumulation (new entrants)

In the old system, a new entrant’s judgment was built up slowly by doing the “form and craft” work that AI now flattens: junior lawyers doing document prep, junior analysts moving data around, junior designers producing execution drafts, junior engineers writing CRUD. Inside this seemingly menial work, new entrants are repeatedly calibrated against reality, and over ten years a traceable track record of judgment forms.

In the AI era this ladder is being weakened, compressed, broken in some industries. Junior roles are taken over by AI; new entrants can no longer assume “do basic work first, build judgment slowly”.

Worse, the new organisational forms (micro-studios, subject alliances, author-led startups) physically exclude new entrants: a micro-studio’s core subjects are busy doing the work, and tuning AI is cheaper than training new entrants; subject alliances need members with their own authorship; author-led startups hire people who already have judgment. None of these new forms have tolerance for new entrants.

But within the closing loop above, a new path opens for new entrants who do have judgment: direct identification by high-taste employers. The concrete shape of this path is dense self-training cycles: use AI as reviewer, do real small tasks, let AI help you mine experience (the method from Issue 02). The essence of judgment is cycle density (frequency of hypothesis, verification, correction), not seniority. A new entrant doing serious AI programming, even just on small scripts, can accumulate more judgment cycles in a year than the old ladder did in ten.

But this path requires the new entrant to start with non-trivial self-drive and judgment from day one. This is a chicken-and-egg problem: no starting point, no dense cycles; no cycles, no developing judgment. The new entrants who can walk this path are typically those with dense cycle training outside formal education: programming since childhood, sustained writing for years, serious competitive hobbies, long-running self-directed projects. The returns to this early accumulation are sharply amplified in the AI era.

To be honest about one thing: the probability distribution of top creators in the future still skews toward survivors with old-era capital and credentials. The new path opens, but the friction hasn’t disappeared: high-taste employers are few, and the self-drive prerequisite keeps most potential new entrants out of dense cycles. The new path is narrower than the old ladder; the share of new entrants who can walk it is lower than in the old ladder era. This is unfair, but it’s the fact.

That said, for young people who choose the new-entrant path, there’s no false pessimism here either, just a different path. The next issue, on education and training, returns to this question.

---

Chapter 8 · What This Means for Individuals: Concrete Actions

Chapter 7 covered the underlying mechanism and the three classes. This chapter gives the most useful concrete actions for the first and second classes (people with accumulation): how to build authorship and relational capital, and how AI amplifies them.

Authorship and relational capital, at the individual scale

The two carriers (node-type / network-type) from Chapter 5, at the individual scale, are called authorship and relational capital. Concretely:

Authorship: recognisable judgment (a clear stance in a field; expression and decisions in different contexts traceable to the same judgment framework), track record (what you did, what you said, the consequences you bore — publicly verifiable), sustained presence (showing up in one field long enough).

Relational capital: long-term collaborators (people you’ve done multiple projects with, who calibrate against you), validated partners (people you’ve gone through hard fights with), deeply trusting clients (5-20 clients willing to hand you the questions that really matter), peer relationships in the field (suppliers, regulators, media — the relationships in which information that AI cannot find flows).

The two operate by different mechanisms: authorship is concentrated in you as a subject — leave, and the value disappears; relational capital is distributed across a set of relationships, but requires you to keep maintaining them. They are not substitutes: authorship lets strangers choose you, relational capital lets deep clients choose you. Most knowledge workers need both.

What the firm used to do for you, you now have to do yourself

This is the most important judgment of the chapter.

Firms used to do two things for you at the same time. They gave you a platform on which to build authorship: the firm’s name card, project opportunities, internal promotion path, letting you accumulate judgment and track record in a stable environment. They gave you colleagues with whom to build relational capital: coworkers, cross-team collaborators, clients, suppliers — relationships the firm reached for you.

Issue 03 covered the disintegration of organisational forms. The structure where the firm “did both for you” is breaking down in the AI era. Micro-studios, subject alliances, independent creators — these new forms don’t do those two things for you anymore.

For most knowledge workers, the most real challenge in the AI era is structurally losing the channels for building these two assets — not “unemployment”.

You may still be in a job, and your salary still arrives every month. But if you aren’t building recognisable authorship in this role, and you aren’t building deep relational capital, then your time is being consumed, not accumulated. Ten years from now when you leave, what you take with you is a job-title resume, not a 10-year moat.

A job-title resume could still be monetised in the old system, because the old filter was formal (large firm brand, job title, project name). In the new system, it has no monetisation channel.

Time is necessary, not sufficient

Before the concrete actions, one meta-judgment: time is necessary, not sufficient.

Authorship needs time: traceable judgment requires sustained public output, track record requires long-term accountability, presence requires staying around. Five years is the floor, ten years counts as initial establishment, twenty years reaches maturity. Most people overestimate their accumulation speed and underestimate the time required.

Relational capital takes longer. It requires both sides to invest time: you can speed up your half, you can’t speed up the other side’s half. This is why relational capital is the scarcest moat: it requires two subjects to bear the cost of building, and most people are only willing to build weak ties.

But time alone isn’t enough. People who only have time discover their time doesn’t monetise: they may have long tenure, but they haven’t built a recognisable subject identity, and they haven’t built a deep relationship network. Their time becomes a sunk cost.

Time is necessary, and you also have to avoid burning time on the wrong track.

The non-accumulation trap

The biggest trap is ten years in, no accumulation. Two typical shapes:

Biological API trap (authorship direction): “biological API” is a concept from Issue 03. Your work is really a human interface to some system (an ERP, a CRM, a workflow), and the work is going to be taken over by AI sooner or later. In a biological-API role, no matter how much time you spend, authorship doesn’t accumulate, because you aren’t making judgments, you’re executing flows. The test: leaving this firm, does your “experience” of the past few years carry value to anyone else? If the value is mainly “familiar with this firm’s internal flows”, you’re a biological API. If the value is “has handled this kind of real problem”, you’re not.

Relationship-instrumentalisation trap (relational capital direction): typical shape is “contact when needed, no contact when not”. This kind of relationship is useful at the weak-tie layer; at the deep-tie layer, it’s the opposite of relational capital: it continuously draws down your trust deposit with the other side without replenishing it. Maintaining a deep relationship requires non-instrumental sustained contact: staying genuinely connected even when you don’t need anything from the other side.

Both traps share one feature: formally looks like accumulation, in substance doesn’t accumulate. Identifying these in the AI era matters more than before, because in the old system “putting in years at a big firm” itself had monetisation power (resume was worth something); after AI peels that layer away, only real accumulation is left.

Concrete actions for building authorship

After avoiding the traps, authorship has clear actions.

One: sustained public output. Public output is the material carrier of authorship: without public output, authorship can’t be traced or recognised. Forms vary: long essays, podcasts, video, Twitter, code repos, open-source projects, public talks, conference appearances. Form doesn’t matter; what matters is sustainedness and traceability. A serious long essay every three years plus daily social-media expression can form a traceable judgment framework; a friend-circle post every three months cannot.

Two: dare to leave a judgment, and bear the consequences of judgment. “This is my judgment, I think X will Y” — this kind of explicit, falsifiable judgment is the core of authorship. Most people’s public output is restating consensus: citing other people’s judgments, summarising industry views, doing “balanced syntheses”. This kind of output doesn’t accumulate authorship, because there’s no judgment framework of yours in it; the reader finishes and doesn’t know “what does this author actually think”. Daring to leave a judgment means daring to be proven wrong afterwards. This is a cost, and a threshold for authorship. People who haven’t borne wrong judgments aren’t recognised as “subjects with judgment”.

Concrete actions for building relational capital

One: keep a few specific people in your field of view long-term. Relational capital is repeatedly working with and being calibrated by a few specific people, not “knowing many people”. 5-20 deep relationships are more valuable than 500 weak ties. Selection criteria: they have real judgment in their field; they’re willing to bear the real consequences of collaboration; their taste can recognise your value. Picking the right people is the first step of relational capital; picking the wrong people turns ten years of investment into sunk cost.

Two: do real things together, bear consequences together. Relational capital doesn’t accumulate in chat; it accumulates in things where consequences are borne together. A real project done together, a loss taken together, a win together — these shared experiences are the core of relational capital. Eating together, attending the same meeting, chatting in the same group — none of those form relational capital.

AI is an amplifier, not a substitute

With authorship and relational capital in place, AI’s role is amplifier.

A subject with authorship used to be limited by their own time. There’s a daily cap on words written, clients seen, problems handled. AI raises this cap systemically; the monetisation efficiency of authorship is amplified.

Concrete shapes:

• Micro-studios: 1-3 core subjects, a small number of part-time collaborators, with execution work outside the core (first drafts, data processing, client communication, standardised output) done by AI. Narrow client base (tens to hundreds of deep relationships), each anchored by authorship, high price per unit, long-term relationships
• Individual alliances: a few independent subjects share infrastructure, distribution, brand backing, while each retains independent authorship. The alliance itself is light; members collaborate rather than employ
• Author-led startups: founder is the subject, the firm is the subject’s amplifier. Product, service, team — all organised around the subject’s judgment. Scale typically 5-50 people; not chasing big scale

One key distinction: this is not “scaling a standardised product”. Standardised products still exist (SaaS, industrial software, financial infrastructure are still here), but a standardised function alone no longer constitutes a long-term premium source — the de-personalised part is exactly what AI does best. The new subject form is the author’s time, amplified by AI into the capacity to serve more specific clients. Each client is still specific, deep, and personalised; only the number of specific clients one subject can simultaneously serve goes from 5 to 50.

The essence of this amplification: AI replaces the amplification work that used to require a team, and the subject keeps doing the judgment. The author is still present; AI handles execution.

Three concrete self-audit questions

Down to concrete actions, three questions to answer honestly:

One: Of my current job, how much would still be with me five years after I leave this firm?

What you take with you = authorship + relational capital. What you don’t take = form assets attached to the firm (job title, process familiarity, internal relationships).

Two: Which moat am I actually running?

Most people will discover: neither is being genuinely run. Time is going into job execution; authorship hasn’t accumulated (no public output, no traceable judgment), relational capital hasn’t accumulated (only internal job-title relationships).

Three: If I leave this firm tomorrow, why would my next client / employer / reader find me, choose me?

If the answer is “headhunter referral”, “the firm’s brand backs me”, “my job title”, you’re in the old system. The monetising power of those things is decaying in the new system.

If the answer is “they’ve read what I publish”, “a collaborator will refer me directly”, “some clients come to me on their own”, you’re in the new system.

The three questions are for honestly seeing your current position, not to manufacture anxiety. Once you see clearly, you can decide the next step.

---

Luck Sits Above All This

After three classes of reader and concrete actions, one warmer thing has to be said. All the judgments, prescriptions, and self-audits above rest on an implicit assumption: that people can change their position by seeing clearly and trying. That assumption is largely correct, but not entirely.

Luck’s weight has been amplified in the AI era. That’s an honest observation.

In the past, luck mattered, but structural paths diluted it: most people could reach the middle by going step-by-step. In the AI era this dilution is gone; who stays and who falls is mostly luck.

Which means: a lot of people will survive in the middle because of luck, and a lot of people will be pushed out from the same position because of luck. The difference can be as specific as “your client’s leadership happened to change”, “your industry happens to have a regulatory buffer”, “your boss happens to dislike AI”.

Attributing your situation entirely to “ability” or “effort” gets less and less accurate in the AI era. Acknowledging luck is for doing the right things while you’re lucky, not pretending you stand on structure; it’s not a reason to give up on effort.

---

Coda · The Deliverable as a By-product of Substance

“The independently evaluable unit” — that has been the economic-structure definition of the deliverable for the past two centuries. It is a form container that lets a subject’s substance detach from the subject and travel independently, priced and consumed by the market on its own.

This structure depended on form and substance being bundled: only because making form required real substance investment was the container visible evidence of substance.

AI, for the first time, lets form be produced independently of substance. The economic meaning of the deliverable as an “independent unit” is failing on a wide front. Form itself is no longer scarce; the container itself no longer needs to be priced independently. The deliverables that still have value must contain substance only the subject can inject: recognisable judgment, relationships embedded in specific operation.

The causal direction has to invert

Before: a subject injects substance in order to make a deliverable. The container is the goal; substance is the means.

Now: a subject sustains operation in reality, accumulates substance, and the deliverable comes out naturally as the expression of that substantive operation. The goal isn’t the container; it’s substance itself.

The deliverable goes from “independently evaluable unit” back to “by-product of a subject’s sustained operation”.

“I want to make a great product”, “I want to write a bestseller”, “I want to make a viral video” — this container-as-goal posture is structurally ineffective in the AI era. Because AI can produce the same external container at lower cost, and you as the creator have no substance to put inside it.

The opposite posture — “I’m working on a specific thing, and some deliverables overflow from it naturally” — is the production posture that holds in the AI era: a blog as the by-product of real research, a product as the by-product of real contact with users, a proposal as the by-product of really handling problems, code as the by-product of really solving problems.

Without underlying substantive operation, all containers are empty shells. In the AI era, the supply of empty form is unlimited; the substance that can be put inside still has to be produced one day at a time by a subject.

---

One Question This Essay Doesn’t Answer

That closes out the judgments and prescriptions on the 5-10 year scale. The coda closes the core moves at that scale.

But there’s a longer-scale question this essay won’t answer, while acknowledging it exists.

All the judgments in this essay come with a time scale.

Today, and the next 5-10 years, the judgment that substance is what tools cannot touch holds. Today’s AI has no continuous self, isn’t bound to consequences, has no concrete presence; what it produces is corpus about substance, not substance itself. All the prescriptions of the previous eight chapters rest on this judgment, and at today’s scale it is solid.

At least for now, we are safe.

But on a longer scale, this judgment shakes. AI not having continuous self is a contingent feature of current architecture, not necessity. AI not being bound to consequences is a deployment choice, not a structural constraint. AI not having embodied presence is engineering reality, not physical impossibility. These three things, on a 20-50 year scale, will most likely change. When AI has continuous identity, is bound to consequences, is present, and does all of these better than humans — where does the human stand?

This essay doesn’t answer it, because today’s understanding can’t. A 12th-century craftsman couldn’t foresee the meaning structure of a 1950s office white-collar worker. A 1950s white-collar worker couldn’t foresee the AI era of 2025. Our generation trying to figure out the human position in 2075 is most likely futile too; that structure has to be groped out by the people of that era inside their new situation.

But one thing can be pointed out.

The recent-modern thing of “deriving meaning from ability” is itself a contingent product of industrialisation and mass education, isomorphic with the bell-curve society. In the thousands of years before that, most people did not live by “winning on ability”. They lived by getting through specific situations with specific others. A 12th-century farmer was outranked on ability by the king, the knight, the monk, the merchant. But he loved his wife, raised his children, buried his parents, and within the small range where he could be present, had meaning, connection, responsibility, joy.

That older meaning system will return in an era where AI fully surpasses humans. It returns because there’s no other choice — not because it’s higher.

Our generation may be the last to derive meaning from “I can do more than others”. The next generation has to relearn a way of living that our great-grandparents still knew and that we have forgotten: not central on the ability dimension, only peripheral; but as specific lives, still complete.

Whether this is pessimistic or optimistic doesn’t depend on the fact itself. It depends on whether you can let go of the obsession that “I must have a position on ability”.

People who can’t let go will suffer through the AI-fully-surpassing-humans process. People who can will find they have returned to a longer human normal.

But that’s decades away.

For today, do today’s work well.

---

References and Sources

The main factual claims in this essay come from the following sources.

On the form migration of printing (history of form migration)

• Dittmar, Jeremiah. “Information Technology and Economic Change: The Impact of the Printing Press.” Quarterly Journal of Economics 126 (2011): book prices fell by about two-thirds
• Febvre, Lucien & Henri-Jean Martin. L’apparition du livre. Paris: Albin Michel, 1958; see also Buringh & van Zanden, “Charting the ‘Rise of the West’: Manuscripts and Printed Books in Europe,” Journal of Economic History (2009): the estimate of European book stock rising from about 30,000 to 20 million
• Erasmus, Desiderius. Adagia (early editions ca. 1500): “Is there anywhere on earth exempt from these swarms of new books”

On the form migration of photography (history of form migration)

• Met Museum, “The Daguerreian Age in France: 1839–55”; Bajac, Quentin. Le ciel et la terre, Musée d’Orsay (2003): the expansion of 19th-century commercial photographic studios
• Gernsheim, Helmut. The Origins of Photography. London: Thames and Hudson, 1982: source of Delaroche’s “From today, painting is dead” (whether the quotation is fully verbatim is disputed, but it accurately reflects how the painters of the time reacted to photography)

On the form migration of industrialisation (history of form migration)

• Ford Motor Company historical archives; Ford Heritage, History of the Assembly Line: Model T chassis assembly time falling from 12.5 hours to 93 minutes; price falling from $850 in 1908 to $260 in 1924

On the collapse of the bell-curve society (Chapter 1)

• Pew Research Center, “The American middle class is losing ground” (2015) and subsequent annual updates; Pew Research Center, “How the American middle class has changed in the past five decades” (2022): middle-class population share and income share data
• Chetty, Raj et al., “The Fading American Dream: Trends in Absolute Income Mobility Since 1940,” Science 356 (2017): 398–406: share of children out-earning their parents — about 90% for the 1940 cohort, about 50% for the 1980 cohort

On RLHF and AI default output converging to the median (Chapter 1)

• Ouyang, Long et al., “Training language models to follow instructions with human feedback” (InstructGPT, OpenAI 2022): InstructGPT rater composition
• Bai, Yuntao et al., “Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback” (Anthropic 2022): RLHF rater sourcing (MTurk workers contributed most of the comparison data, Upwork workers a smaller share)
• Durmus, Esin et al., “Towards Measuring the Representation of Subjective Global Opinions in Language Models” (Anthropic 2023, GlobalOpinionQA): aligned models’ bias on subjective issues

On substance as credence good and the quality identification problem (Chapter 3)

• Nelson, Phillip. “Information and Consumer Behavior.” Journal of Political Economy 78 (1970): 311–329: search goods / experience goods
• Darby, Michael R. & Edi Karni. “Free Competition and the Optimal Amount of Fraud.” Journal of Law and Economics 16 (1973): 67–88: credence goods
• Akerlof, George A. “The Market for ‘Lemons’: Quality Uncertainty and the Market Mechanism.” Quarterly Journal of Economics 84 (1970): 488–500: equilibrium analysis of markets where quality is unobservable. Akerlof, Spence, and Stiglitz shared the 2001 Nobel Memorial Prize in Economic Sciences

On signalling theory and mimicry cost (Chapters 3, 4)

• Spence, Michael. “Job Market Signaling.” Quarterly Journal of Economics 87 (1973): 355–374: the core mechanism by which signals sustain market separation is the difference in mimicry cost
• Feltovich, Nick, Rick Harbaugh & Ted To. “Too Cool for School? Signalling and Countersignalling.” RAND Journal of Economics 33 (2002): 630–649: counter-signalling model
• Bellezza, Silvia. “Distance and Alternative Signals of Status: A Unifying Framework.” Journal of Consumer Research (2023): unifying framework for distance-based alternative status signals

On authorship as a modern category (Chapter 4)

• Foucault, Michel. “Qu’est-ce qu’un auteur?” Bulletin de la Société française de Philosophie (1969); English: “What Is an Author?” in Aesthetics, Method, and Epistemology (1998): the formation of the author as a modern category after print

On AI’s complement effect outweighing the substitute effect (Chapter 8)

• Mäkelä, Elina & Fabian Stephany. “Complement or substitute? How AI increases the demand for human skills” (arXiv:2412.19754, 2024; v3 February 2025): based on 12 million US job postings 2018-2023, showing AI’s complement effect can be up to 50% larger than its substitute effect in the data; data scientists with resilience or ethics skills can earn a 5-10% wage premium

Conceptual continuity with previous Offbook Press issues

• “Cognitive decoupling” originates in Issue 01, On Cognitive Decoupling
• The “AI review loop” method originates in Issue 02, Rebuilding Learning
• “Disintegration of organisations” and “biological API” originate in Issue 03, Breakdown of Firms

---

This essay paraphrases the sources conceptually rather than quoting them directly. For verification, please refer to the original publications.

---