Abstract　The impact of generative artificial intelligence on distributive order is usually framed as a problem of unemployment. This article argues that the order is reversed: what fails first is society’s capacity for screening, that is, its ability to determine at low cost what any given person can do; the reorganisation of the labour market follows from that failure rather than preceding it. Drawing on ten years of interaction data from the ALEKS platform, a survey conducted by the China Youth and Children Research Centre, US administrative payroll records, and several 2026 measurements of model capability and compute distribution, the article works through the gap in existing research on the screening variable, the three boundaries of substitution, the way material conditions set a price floor and a rent destination, the direction in which labour power has moved, the historical forms in which societies have handled surplus populations, and the manner in which the centre of value shifts with scarcity, before setting out observable checkpoints. The core claims are three: rising productivity only extends the boundary of the possible and does not select where society lands; when the marginal cost of cognition approaches zero while the rate of material conversion does not, rent retreats to energy and land; and distributive outcomes are determined by political rather than technical processes.

I. Statement of the Problem

A study published in 2026 designed a quasi-experiment capable of isolating the true effect of generative artificial intelligence on learning. The researchers obtained ten years of interaction data from the ALEKS mathematics platform, comprising 3.2 million learning records covering everything from fifth-grade arithmetic to college algebra. Problems on the platform fall into two categories: plain word problems, whose entire text can be transcribed directly into a conversational model’s input, and graphical interaction problems, which require dragging and reading figures within the interface and cannot be reproduced as text. The difference between the two categories constitutes a natural control.

The results show that after the release of ChatGPT, study time on word problems fell by 2.8 per cent per quarter at the university level, a cumulative decline of 26.9 per cent over eleven quarters; 31.3 per cent cumulatively at the high-school level; 9 per cent at the middle-school level; and no detectable change in fifth grade. Study time on graphical problems showed no change. Under proctored conditions, the decline in word-problem study time disappeared entirely. The most discriminating result is the last: on randomly assigned items retained for proctored testing, the odds of a correct answer fell by 25 per cent cumulatively, while the same estimator applied to unproctored tests yields an 85 per cent increase. The same students perform better when unobserved and worse when observed. The researchers named the phenomenon cognitive surrender, distinguishing it from cognitive offloading: when a calculator performs a multiplication, only the arithmetic is offloaded while the structure of the solution remains with the solver; when a problem is submitted to a conversational system and its output adopted, what is surrendered is the reasoning itself.

Common readings of this result focus on teaching methods and academic integrity. This article argues that the more consequential issue is one of screening mechanisms. The social force of degrees, examinations and rankings rests on the premise of measurability; once measurement fails, every arrangement built upon it is implicated. This can be stated as a more general proposition: distributive order depends on the low-cost verification of ability, and technological change may dismantle that capacity for verification without altering total employment. This article therefore treats the failure of screening as an independent variable and reorders the sections accordingly, taking measurement and substitution first and distribution and belief afterwards.

II. Existing Research and Its Limits

(i) Two Positions in the Literature on Technological Unemployment

Research on the relationship between technological change and employment divides roughly into two positions. The compensation school holds that jobs destroyed by technology are offset by new ones. This position can be traced to the 1966 report Technology and the American Economy, which answered the 1964 “Triple Revolution” memorandum and its claim that automation would end employment; in every subsequent wave of automation anxiety the compensation argument has been invoked again, and in terms of total employment it has repeatedly been vindicated.

The other position holds that the present transition differs in kind. Frey and Osborne, assessing the automatability of 702 occupations in 2013, concluded that roughly 47 per cent of US employment lay in the high-risk range. Acemoglu and Restrepo distinguished the displacement effect from the productivity effect in 2019, arguing that the net effect of automation on the labour share depends on the relative strength of the two. Brynjolfsson proposed the notion of the Turing Trap in 2022, warning that a technological trajectory aimed at substitution suppresses both labour demand and the wage share.

The disagreement between the two positions turns on the dependent variable: the compensation school measures total employment, the change-of-kind school measures factor distribution, and the same data can therefore support opposite conclusions. This very disagreement suggests that distribution and employment need to be handled as separate variables, which is the starting point of the framework set out in Section III.

(ii) Why Screening Has Not Been Treated as a Variable

A second strand of research relevant here comes from the economics of education as signalling. Spence argued in 1973 that the primary function of education is not to raise productivity but to transmit a signal of ability to employers, whose willingness to pay depends on the credibility of the signal rather than the content of the knowledge acquired. The theory has since been used extensively to explain the signalling component of returns to education.

Signalling theory and its empirical literature, however, presuppose that the signal is stable; they do not address the case in which the signal itself fails. Generative artificial intelligence reduces the measurability of ability, and systematic evidence on this point has appeared only since 2024: apart from the ALEKS study cited above, it includes work at Stanford and the Dallas Fed on entry-level employment, along with a number of experiments on performance differences under supervised and unsupervised conditions. Existing research has noticed these phenomena but has mostly filed them under labour-market disruption or educational technology; it has not treated the failure of screening as an independent variable whose relation to distributive order is worth examining. That gap is where this article enters.

III. Analytical Framework

The analysis proceeds at three levels, defined here and then used without restatement.

The first is the three tiers of substitutability. At the technical tier, any task with a well-defined objective and automatically verifiable outcome can be substituted. At the economic tier, the actual speed of substitution depends on the comparison between the marginal cost of running artificial intelligence and the cost of human labour; artificial intelligence consumes electricity, chips, depreciation, land and cooling water, so the price floor of human labour is set by physics rather than by ability. At the institutional tier, liability, licensing and the legal authority to sign constitute constraints: medical diagnosis, legal opinions, audit signatures, safety certification and airworthiness standards impose requirements of human responsibility that are written into statute rather than into technical capability. Conflating the three tiers is the main reason existing discussions reach contradictory conclusions.

The second is the four types of distributive mechanism. Once labour ceases to be the channel of distribution, distribution must occur by other means, and history offers only four: asset ownership, political allocation, violent appropriation, and reciprocity within households and communities. Which one prevails is decided politically. Determining which type a given historical arrangement belongs to clarifies both its stability and its cost.

The third is material conditions as a constraint. The marginal cost of cognition can approach zero; the rate of energy conversion and the area of land cannot. Distributive struggle therefore has a budget line, and the struggle itself cannot take place outside that boundary. The same constraint determines where rent lands: when the cost of copying approaches zero, value retreats to what cannot be copied.

IV. The Failure of Screening

(i) Two Sets of Evidence

A 2025 survey by the China Youth and Children Research Centre covered seven provinces, urban and rural, and collected 8,563 valid responses. More than sixty per cent of primary and secondary school students had used artificial intelligence; the most common purpose was assistance with homework, at 71 per cent. Nearly seventy per cent of households had no rules governing such use, and the share of rural parents who exercised no supervision was 20.3 per cent, against 15.6 per cent in cities. More than twenty per cent of students said they preferred not to think for themselves, and over ten per cent trusted generated content entirely. Nearly half of the children turned to artificial intelligence rather than to people around them when something troubled them, and more than twenty per cent preferred talking only to machines. The share of rural students using artificial intelligence to write assignments on their behalf was 18.8 per cent, against 14.5 per cent in cities. The scholar who led the study relayed a frontline teacher’s perplexity: it is no longer possible to judge whether a student’s homework was written by artificial intelligence.

Survey data alone do not establish a failure of measurement; what the survey records is usage, not the capacity to discriminate. Placing that perplexity alongside the ALEKS study cited in Section I yields a mutually corroborating picture: the first is the failure of discrimination as practised, the second is the same phenomenon projected onto large-scale behavioural data, and the second supplies a discriminating piece of evidence the first lacks, namely that the gap vanishes entirely under proctoring. The two sets of material differ in period, sample and object of measurement, yet point in the same direction.

(ii) Why the Failure Is Stratified

The failure is not universal but stratified by type of ability. Capacities requiring real operation, on-site judgement and long-term tracking remain measurable, though at rising cost; capacities that depend purely on text and information processing are the easiest to counterfeit and were, for that same reason, the easiest to measure at scale, and over the past two centuries it was precisely this portion of ability that carried the main burden of social selection. The stratification has a structural cause: codifiable knowledge can be automated and therefore counterfeited automatically, so the difficulty of verifying it rises in step with the difficulty of generating it. Tacit knowledge and on-site skill behave in the opposite way: verifying them requires the observer to possess comparable experience, which makes verification hard to scale but also hard to fake. The cost structure of screening thus inverts: what was once the cheapest and most measurable becomes the least trustworthy, while what was once expensive and hard to measure remains credible.

V. The Boundaries of Substitution and Material Conditions

(i) Where the Gains Went

Technical substitutability does not translate automatically into economic substitution, and the distributive effect is the part most easily overlooked, though it happens to have the hardest quantitative evidence. Between 1855 and 2020 the labour share of income in major economies remained stable at around two-thirds. Researchers at the Federal Reserve Bank of Philadelphia explain this as a random balance between two effects: displacement, which shifts tasks from labour to machines and depresses the share, and productivity, which raises the efficiency of machines on already-automated tasks and lifts it. They further observe that if the new technology systematically favours the displacement of tasks not yet automated, that balance may break and the labour share will decline as a trend, even with total employment unchanged. One qualification is necessary: the evidence for this judgement comes either from the period before generative artificial intelligence, that is, the industrial automation of 1980 to 2016, or from a window of only about three years, so the experience of older technologies can illustrate a mechanism but cannot be extrapolated directly into a trend.

Acemoglu and Restrepo provide another set of estimates. They find that automation endogenously targets high-rent jobs, meaning positions whose wages exceed workers’ opportunity costs, and that this accounts for 52 per cent of the rise in between-group wage inequality in the United States since 1980. Such targeting is also inefficient: it offsets between sixty and ninety per cent of the productivity gains from automation, and between 1980 and 2016 automation raised total factor productivity by a net 0.3 to 1.3 per cent. The dependent variables differ from those of the labour-share research, inequality in one case and factor distribution in the other, but both support a single inference: the gains from technological progress are largely consumed by distributive struggle before they reach society.

(ii) The Generation That Cannot Get In

When the contemporary form of this reorganisation appears in employment data, its shape runs against intuition. Using payroll data covering 25 million workers from ADP, researchers at Stanford found that, after controlling for firm-level shocks, relative employment among 22 to 25 year olds in the most AI-exposed occupations fell by 16 per cent, with software development down nearly twenty per cent from its peak at the end of 2022, while employment for experienced workers held steady. The Dallas Fed replicated the finding on a different dataset, obtaining a 13 per cent decline and isolating the mechanism: what drives the fall is a reduction in the inflow of workers entering employment from outside the labour market, not layoffs, and its contribution to the overall unemployment rate is 0.1 percentage points. The corresponding figures for China are an urban youth unemployment rate of 18.9 per cent for the 16 to 24 age group in August 2026, the highest of the year, and 12.7 million university graduates entering the market in the same year, more than sixty per cent of whom settle in prefecture-level cities.

These data point to a specific shape: the young cannot enter, and the ladder from apprentice to expert has been withdrawn. Compared with mass unemployment this shape is harder to see, because it does not appear in aggregate indicators but only in intergenerational mobility, and the latter is not published monthly. It also has a rarely mentioned consequence: if artificial intelligence substitutes for the codifiable knowledge of the entry stage while preserving the tacit knowledge that requires accumulated experience, then twenty years from now there will be a shortfall in the supply of experts, since experts were previously produced by entry-level work. The inference depends on substitution genuinely being stratified by codifiability; if technical development makes tacit knowledge equally codifiable, the shortfall will not appear.

(iii) The Actual Distribution of Compute and Models

Several measurements from September 2026 yield a more detailed picture than the phrase “open against closed” suggests. At the top of the stack, compute and energy are highly concentrated: six companies account for seventy to seventy-five per cent of global hyperscale data-centre electricity, and more than ninety per cent of compute capacity sits in North America, Western Europe and Asia-Pacific. At the model tier, since January 2026 the strongest open-weight models have lagged the frontier closed models by an average of four months; broken down by country, Chinese open models trail the US closed frontier by roughly two to five months, while US open models trail their own country’s closed frontier by six to nine months. Data on adoption are more telling still: cumulative downloads of Chinese models on HuggingFace are about twice those of US models; on OpenRouter the share of calls going to Chinese models has risen from about seventy per cent to over eighty, with total volume growing from one trillion tokens a week to eighty trillion within a year; more than ninety-five per cent of inference on the open-source coding agent OpenCode goes to Chinese models; and in the main arXiv machine-learning categories, more than forty per cent of papers mention a Chinese open model, against about thirty per cent for US models. Distillation explains only one or two months of the gap. At the tier of distribution, compliance and trust, closed providers retain the advantage, resting on compliance documentation, enterprise procurement inertia and liability clauses, all institutional factors.

The technical report released by DeepSeek in September 2026 supplies a concrete cross-section. The model carries 552 billion backbone parameters and 196 billion Engram parameters while activating only 8 to 16 billion per token; its global KV cache is compressed to 890 bytes per token, one quarter of the previous generation and one four-hundred-and-thirty-seventh of the first; context length extends from 4K to 1M, a factor of 256, while decode compute rises by only about a quarter; its base model matches the previous generation’s flagship with one third of the total parameters, and the post-training stage contains no algorithmic innovation at all, every improvement coming from the data pipeline. The figures point to an explicit trajectory, namely the separation of capability from scale. They also answer a common dispute about whether China’s efficiency advantage stems from state subsidy or from constrained compute: subsidy typically manifests as more hardware, whereas every improvement here targets output per unit of compute, which is closer to an engineering choice made under constraint.

(iv) Electricity and Land

If the marginal cost of cognition can keep falling, the limiting factor ceases to be the algorithm and becomes those links the algorithm cannot reach. Artificial intelligence is itself material-intensive: data-centre electricity consumption is expected to grow from about 118 TWh in 2024 to 239–295 TWh by 2030, roughly one per cent of global electricity. The tighter constraint lies in grid interconnection: for projects completed in 2025, the median time from interconnection request to commercial operation was 61 months, against 36 months in 2015 and 22 months in 2008; of the 2,290 GW in the queue in 2024, only 53 GW came online in 2025, a throughput of 2 per cent, while 600 GW of new requests were filed and 750 GW withdrawn in the same year.

The pace of physical construction does not answer to capital, and what is genuinely scarce is transformers, transmission corridors and permits. Between 2025 and 2026, the principal means of relieving this bottleneck was to use artificial intelligence to accelerate the permitting process itself; one grid operator’s pilot cut study time by 98 per cent. This marks the boundary of what artificial intelligence can do: it reduces institutional friction, not physical constraint. The observation also qualifies the earlier judgement about substitution.

VI. The Transfer of Labour Power and the Historical Forms of Distribution

(i) Power That Came from Presence

Before discussing distribution it is necessary to see the structure of labour’s power. Political sociology distinguishes two kinds of worker power: structural power, the capacity to withdraw labour that derives from a position in production, and associational power, the organisational capacity of unions and parties. The distinction was set out systematically by the American sociologist Erik Olin Wright in his analysis of class. The power of the labour movement derived chiefly from capital’s dependence on workers being present, never principally from numbers. Fordism at mid-century was the high point of structural power: production was finely divided, skills were specific, halting output was extremely costly, and strikes therefore acted directly on capital with visible effect.

Three mechanisms have since weakened it in concert: capital can relocate plants, substitute machines, and arbitrage through migration and outsourcing. The trend has run from the 1950s to the present, showing up as a long decline in union density, a weakening of strike effects and a falling labour share. The process is one of narrowing rather than disappearance: strike waves in Britain and the United States in recent years, among screenwriters, auto workers, health care and railways, show that a minority of irreplaceable positions retains structural power while the majority has lost it. This explains two simultaneous phenomena: premiums rising for key occupations such as nurses, electricians and lorry drivers, and the bargaining capacity of ordinary white-collar workers falling. The total quantity of power need not have diminished; its distribution has polarised sharply.

(ii) Power That Comes from Location

Weakening and compensation occur together. Industrial capital can move globally, which is the main reason workers’ structural power was undermined; the means of production for artificial intelligence, by contrast, cannot be moved, since grid capacity, water, land and planning permission are all fixed to particular places. The consequence is that host communities acquire a veto, and that power is already in operation: a data centre in County Mayo, Ireland, had its permission overturned on environmental grounds; an opposition movement formed in County Kildare, with two hundred people attending a meeting in Naas and protesting outside parliament; a project opposed by the writer Sally Rooney had its planning consent overturned on climate grounds; and cancellations are accumulating in the state of Virginia. The headline of one European report reads plainly: power-hungry data centres are expanding, and so is the resistance.

These struggles do not take the form of strikes but of planning permission, environmental impact assessment, electricity price negotiation and local elections. Their character is political and legal, and the threshold they require is location rather than numbers of workers. The field of conflict has therefore shifted from the factory floor to the sites where compute is installed. The shift has theoretical significance: it moves the point of entry into struggle out of the employment relation and into the systems of space and permission, which are open to all residents and do not require being hired first. The earlier judgement about the collapse of structural power therefore needs qualification: what collapsed was the kind of power carried by the employment relation, while power grounded in location acquired a new object in the same period.

(iii) Four Precedents

Modern history offers four precedents, which are not unrelated to one another but converge on a single conclusion.

In 1795, magistrates in Berkshire, England, decided that when a labourer’s income fell short of subsistence the parish would make up the difference. This was the first large-scale wage subsidy in history, known as the Speenhamland system, and its consequences were subtle: because the parish bore the difference, employers had an incentive to hold wages down, so labourers gained subsistence while losing bargaining power. Thirty-nine years later the New Poor Law of 1834 abolished outdoor relief, built workhouses, and established the principle that relief inside the workhouse must be worse than the lowest wage outside it. The modern counterpart of this sequence is nearly isomorphic: income guarantees correspond to Speenhamland, and workfare corresponds to the New Poor Law. The lesson is that when demand for labour is insufficient, the first response of society is subsidy and the second is punishment, and a decent arrangement does not appear of its own accord but must be won through political pressure.

In March 1964 a group of American intellectuals published the “Triple Revolution” memorandum, claiming that automation would end employment and calling for a guaranteed income. Two years later the United States established the National Commission on Technology, Automation, and Economic Progress, which in its 1966 report rejected the claim, arguing that destroyed jobs would be offset by new ones. On total employment the commission was right, but its account of the mechanism was wrong: what was displaced then was manual and clerical labour, while the demand that simultaneously erupted was for cognitive labour, in management, engineering, education, medicine and finance. This is where historical analogy most needs caution, since the vehicle of the last transition is precisely the object of this one’s substitution. The sceptics of the 1960s were right in their conclusion, but the grounds on which they were right no longer hold. Another fact that should not be overlooked is that after 1966 unemployment did not rise markedly; inequality did, which matches many of today’s predictions.

The form of compulsory employment can be illustrated by Soviet full employment and by the lifetime employment of Chinese state enterprises together, since both ended at the same point. Full employment as a political commitment produced concealed unemployment: overstaffed enterprises, proliferating formalistic posts, low efficiency, everyone nominally employed and no one actually needed. The Chinese counterpart ended with the layoff wave of the late 1990s, on a scale of tens of millions, at enormous one-off political cost. The historical forms of compulsory employment are thus only two: establishment expansion, which has a fiscal ceiling, and formalistic posts, which sacrifice both efficiency and dignity. Late Roman grain distribution and public games supply a fourth case: when labour is no longer needed, the durable stable form is to be provided for, to be entertained, and to remain politically indifferent, which structurally resembles the contemporary combination of income guarantees, algorithmic recommendation, programmatic companionship and the games industry. That form is stable, and its stability rests on a core capable of extracting external resources: the provinces then, the concentrated owners of compute and energy today.

(iv) Why Organisational Capacity Becomes Decisive

Taken together the four precedents establish two things. First, a rising share of the political does not mean a rising share of the democratic; the historical norm is a narrowing of the participants alongside an expansion of the nominal scale of distribution. Second, expecting honest government and peasant rebellion do not exhaust the options: between them lies a third path, organised political participation, and whether that path exists depends on organisational capacity.

On the first point, the establishment of welfare states has always been accompanied by struggle: Bismarck’s social insurance had a clear motive of pacification, the British National Health Service was the work of the post-war Labour Party, and the American New Deal was the joint product of the Depression and a strong union movement. Every improvement in distribution has had an organisation behind it. On the second point, the analogy of peasant rebellion has its preconditions, requiring subsistence crisis, collapse of state capacity, and an alternative narrative of legitimacy all at once, and modern conditions do not meet them. The multi-country movements of 2011, Chile and France in 2019, and the United States in 2020 are not rebellions but decentralised, periodic eruptions whose common feature is the capacity to overturn particular policies without the capacity to establish an alternative order. The reason is the absence of organisation, and that absence is itself a by-product of declining structural power.

VII. The Displacement of the Centre of Value

(i) From Agrarian to Industrial Society

Beyond interests and power there is a further layer that productivity changes, more slowly and more deeply: the system of social value, that is, what is esteemed, what is disdained, and where identity and dignity come from. Understanding this layer requires first acknowledging how it moves: value gathers around whatever is scarce, and the process always lags.

Agrarian value revolved around land and risk: diligence, thrift, obedience to elders, many children. This set of values was shaped by the mode of production, with moral philosophy supplying only retrospective explanation: agriculture is seasonal, exposed to crop failure and dependent on long-term investment, so thrift and diligence were survival strategies and the family was an insurance unit. Industrial society reshaped those values: the factory required punctuality, discipline, specialisation and replaceability, so diligence turned from a survival strategy into a moral obligation. The mechanism Weber described in The Protestant Ethic and the Spirit of Capitalism is the crucial link: Calvinism interpreted the secular calling as a vocation, and profit-making turned from a suspect activity into moral proof. Value here supplied legitimacy to a new mode of production.

But the old values did not withdraw at once. Veblen described another order in The Theory of the Leisure Class in 1899: the propertied established status by abstaining from productive labour, conspicuous consumption and conspicuous leisure coexisting. While the cult of diligence became mainstream, another class continued to mark identity by leisure, two standards of evaluation existing at once and in conflict. The juxtaposition indicates that the value structure of a transition period is usually not a single replacement but a competition among multiple standards, whose outcome depends on which class holds discursive power at the time.

(ii) What the Service Economy Actually Delivered

The service economy’s experience warns that a so-called upgrading of industrial structure does not guarantee an improvement in the value system. Daniel Bell predicted in The Coming of Post-Industrial Society in 1973 that services would dominate and knowledge would become the core resource. The prediction was correct as to industrial structure, with manufacturing employment falling from a quarter to under a tenth and services reaching eighty per cent; but the direction of value change diverged from common expectations.

The sociologist Arlie Hochschild, studying flight attendants in 1983, introduced the concept of emotional labour: the job requires workers to manage their own emotional expression, with smiling, patience and empathy becoming components of the labour process subject to assessment. This produces a paradox in which the most human element is demanded in the most mechanical way. The anthropologist David Graeber recorded another phenomenon in Bullshit Jobs in 2018: a great many jobs are regarded as meaningless even by those who hold them, yet they continue to be created. His explanation is that the actual function of such work is distribution: where productivity no longer requires the labour, work still serves to deliver income and identity. The insight reaches beyond its surface: the social function of work long ago separated from its productive function, which explains why, as productivity rose, humanity did not reduce working hours but invented more positions.

At the same time the labour market polarised, with intermediate manufacturing and clerical jobs disappearing and both high-skill and low-skill services growing. The service society is therefore not a homogeneous whole but contains two radically different situations, professional work in finance, consulting, software and medicine at one end, and monitored, poorly regarded, low-mobility work in food service, retail, care and delivery at the other, and the split in the value system occurs along this fault line. The polarisation also indicates that a single scale misleads when discussing value change, since two opposing value movements may coexist in one society.

(iii) Four Regularities and Their Limits

Placing these histories side by side reveals several relatively stable regularities. Value gathers around scarcity: when something becomes cheap, the esteem surrounding it declines. The value system lags: the values of an old mode of production continue to govern for a considerable time under the new one, and a cohort invests according to the old rules only to be settled by the new. This one has a fairly precise historical measurement, since the economic historian Robert Allen found that between the 1790s and the 1840s per capita output grew continuously while real wages stagnated for nearly half a century, a phenomenon later called the Engels pause, during which that generation of workers stood between two sets of rules. Old values do not disappear but change form: land was identity in agrarian society, property became identity through industrialisation, and assets or access to compute may become identity in future. And transitions always involve two sets of values existing simultaneously and in conflict, since the depth of social conflict is often not merely a conflict of interests but a disagreement about what counts as a good life.

From this one can reason about conditions in which production becomes cheap. Work supplies three things: income, identity and temporal structure. Income can be replaced by transfers, temporal structure by other arrangements, and identity is the hardest to replace; the source of identity may therefore leave occupation for community, interest groups, religion and local belonging, which is why cultural questions carry a rising share of politics. A possible reversal also deserves note: in the framework of The Theory of the Leisure Class, both leisure and work are instruments of status, the difference lying in who is entitled to leisure; if work ceases to be necessary, being able to work may become a privilege, and there are already signs of it, with those replaced by automation envying those still employed and office workers envying the bodily freedom of manual labour. Related to this, the status of care work may be reassessed: historically, raising children, nursing and housework were treated as non-labour, excluded from gross domestic product and unpaid, a statistical system criticised systematically by the feminist economist Marilyn Waring in If Women Counted in 1988; if artificial intelligence makes production cheap, the labour closest to people may instead become the centre of value.

Several qualifications must be retained. The value system does not follow the mode of production automatically: it has carriers in the family, the school, religion and the media, each with its own interests and inertia, and after a technological change it may lag for decades or be deliberately maintained by interest groups. Form is predictable while content is not: the logic of scarcity indicates the direction in which value gathers but cannot say what people will specifically esteem, and historians have a poor record of predicting changes in social value. Expectations that industrial upgrading brings humane outcomes require caution, since the service society has already supplied a counterexample. And if material constraints tighten, the value system may regress: should pressure on energy, ecology and food rise, survival values will reassert themselves. Beyond the general regularities, China’s recent decades offer a concrete case: the core value of the reform era was prosperity through diligence, corresponding to rapid growth and relatively open channels of advancement; the term involution then appeared, describing intensified competition with stagnant returns; and after that came the rising popularity of civil service examinations and establishment posts, a shift from seeking opportunity to seeking certainty. The three terms appeared in succession, corresponding to changes in the width of the channel, which indicates that value shifts are often responses to structure and that the response lags.

VIII. Discussion: People in Different Positions

(i) Ordinary Lives

The mechanisms above ultimately settle into concrete situations, and the differences between positions are far greater than averages suggest. For most people the price structure will split: what can be digitised tends toward zero marginal cost while what cannot becomes more expensive, including care, repair, education, medicine, catering and construction, so household budgets will show two extremes, with entertainment nearly free and care expensive enough to require years of saving. At the same time artificial intelligence consumes electricity and land: data centres are pushing up local electricity prices and water stress, and drawing electricians and HVAC technicians away from residential construction.

The change in employment form appears as delayed adulthood rather than mass unemployment, with young people returning to their parents’ homes, continuing to study, preparing for civil service examinations and repeating internships. This change will not appear in unemployment rates, only in the sequencing of lives. The relative purchasing power of assets and labour will also shift: a falling labour share means falling relative purchasing power, and for goods in fixed supply such as housing and land, prices are set by average ability to pay, so housing may become less affordable even as wages rise, which explains the persistent gap between aggregate growth narratives and individual experience.

(ii) The Post-2010 Generation

Those born between 2010 and 2019 hold particular research interest: in 2026 the oldest are sixteen, and they will enter society between 2028 and 2038, having grown up entirely after the diffusion of generative artificial intelligence, which makes them a natural sample for testing the judgements above. They differ from the previous generation in three respects. Artificial intelligence is part of their environment rather than a skill for a résumé; the previous generation regarded fluency with it as an advantage, whereas for them it is a default condition, as nobody today regards the ability to use a search engine as an advantage, and familiarity confers no competitive edge. The education they receive is losing its measuring capacity: homework cannot be reliably distinguished as genuine, unsupervised assessments are inflated, retention under supervision declines. This means the chain between effort and reward is broken: those who think seriously go unrecognised in an unsupervised system of evaluation, while those who take shortcuts obtain good marks. And they enter the market precisely in the interval when the entrance is closing, the previous generation having entered during expansion and theirs during contraction, a rupture that stems from the year of birth and has little to do with individual quality.

China has an additional line of division: rural students use artificial intelligence to write assignments at a higher rate than urban students, and rural parents exercise no supervision at a higher rate as well. Using artificial intelligence as a learning scaffold requires parental guidance, so the technology is likely to widen rather than narrow educational inequality. A deeper consequence is visible in the figure showing that nearly half of children turn to artificial intelligence first when something troubles them: their emotional regulation and social capacities are formed in interaction with an object that is always patient and always accommodating. These characteristics are inferences, and their verification depends on longitudinal data over the next ten to twenty years.

(iii) Choice of Field

Judgements about fields of study should not rest on popularity. Graduates in computer science in the United States face employment prospects resembling a recession, with entry-level hiring falling more than in any field but one; in China the 2026 green-list fields were all engineering, automation entering for the first time, while information security, network engineering and other computer-related fields left the list collectively, and close to 700,000 computer science graduates entered the market that year. Betting on popular fields is always a lagging strategy, since a four-year training cycle will be out of phase with policy and market cycles.

Three criteria are more reliable than rankings. Whether the position carries legal liability: in law, clinical medicine, auditing and safety engineering, responsibility cannot be outsourced to a machine because the signatory must be a legal person. Whether the position requires physical presence: nursing, dentistry, veterinary medicine, repair, construction and laboratory work depend on bodily presence, and demand will rise with population ageing. Whether the position complements or competes with artificial intelligence: in control and automation, energy and power, materials, bioinformatics, electromechanics and hardware, the effect of using artificial intelligence is to amplify output, whereas positions dense in codifiable knowledge that carry no responsibility and require no presence run the highest risk, including some entry-level programming posts, junior legal and financial work, translation, graphic design and media. A field is not an occupation: the criteria concern the position, not the degree.

(iv) The Abilities That Intelligence Tests Measure

Those in the top one per cent of their cohort by cognitive ability occupy a different situation, and this is the section requiring the most care. The abilities measured by intelligence tests, including abstract reasoning, pattern recognition, working memory and processing speed, are precisely those generative artificial intelligence takes over most readily. The mechanism described in the Stanford study is more specific: artificial intelligence substitutes for codifiable knowledge and preserves tacit knowledge, meaning the skills accumulated from experience that have never been digitised, and tacit knowledge correlates weakly with intelligence while correlating more strongly with experience, presence and environment.

More importantly, the chain of screening has broken. The historical pricing of high intelligence rested on three conditions: scarcity, verifiability and leverage. Artificial intelligence has shaken the first two at once, since cognitive execution is commodified and ability signals can be forged where supervision is absent; when ability cannot be verified, the scarcity premium built on measurement collapses. Breaking cognitive activity down, the order of substitution runs roughly as follows: reasoning and computation to execute a known task are almost entirely substituted and correlate most strongly with intelligence; optimisation within a given framework is largely substituted; defining the problem is only partly substituted and correlates moderately; judging what counts as good depends on taste and correlates weakly; bearing responsibility cannot be substituted; physical presence cannot be substituted.

The conclusion should therefore be put this way: high intelligence has not ceased to matter; what has changed is how it is priced, from a scarcity premium to a premium on choosing and bearing. The former could be obtained through intelligence alone; the latter requires elements beyond it, namely taste, long-term commitment, networks of trust and access to capital. This has happened before: literacy, imperial examinations and degrees were all once marks of elite status and later became baseline conditions, and artificial intelligence may make intelligence a baseline condition too.

One specific configuration deserves separate examination. Its features are cognitive ability confirmed through systematic selection and already detached from the standard track, use of artificial intelligence for production rather than for completing assignments, and the existence of a locally validated path model. What tends to happen to such people is not devaluation but early entry into a new pricing system: they are already making choices and answering for them. Yet four points warrant sobriety: behind the visible successes are silent failures, and the retreat costs of the competition path are high; the sense of cognitive leaps produced by a training camp usually exceeds the measurable gain in ability and requires external benchmarks to calibrate; intensity of input is not quality of output, and the criterion is always what has been made; and an early-stage track means both first-mover advantage and unsettled institutions. For them and for all those with marked cognitive ability, one judgement applies: technical leadership is not the same as institutional security. The real question is whether technical ability can be converted, between the ages of twenty and thirty, into some form of ownership, be it verified output, an irreplaceable position, a replicable system, or access to capital; the most dangerous configuration is pronounced ability without assets, since in an age dense with capital, ability can be rented and rent accrues to owners.

IX. Conclusion

Scenario Content Trigger Indicators
Rentier society A few owners control interfaces and infrastructure while most people live on transfers and dependent services Rising concentration, falling labour share, rent destination fixed AI rent as a share of GDP, compute concentration, model accessibility
State-led socialisation Compute treated as infrastructure, returns channelled back through taxation and factor distribution Security and industrial competition overriding pure market logic Public compute share, data factor distribution rules, platform bargaining mechanisms
Social democracy Taxation of compute, data and rent, paired with income guarantees, shorter hours and public services Labour and consumers organised enough to form a political majority Actual working hours, whether income guarantees shift from consumption subsidy to asset sharing, remuneration of care work
Stagnation and paradox Technology arrives without growth, artificial intelligence reduced to a cost-cutting tool No acceleration in total factor productivity, investment bubble bursting Total factor productivity, unit labour cost, rent-dissipation ratio

The argument reduces to eight observable judgements. The narrowing of structural power should appear as rising premiums for key occupations together with a falling bargaining capacity among the general workforce. The localisation of conflict should appear as struggles over permits, environmental review, electricity prices and local elections rather than on the factory floor. Organisational capacity should replace structural position as the decisive variable, with the ability to organise mattering more than numbers. The form of social conflict is more likely to be periodic, decentralised unrest than rebellion, destructive without being constructive. Coercive apparatus should become the last safe harbour, observable as counter-cyclical employment growth in policing, the military and security. If the state holds ownership, distribution will be politicised and relatively controllable; if the state is captured, debt, inflation and polarisation will arrive together. The rate at which unit intelligence cost falls will predict the destination of rent better than the capability gap between open and closed models: if open models keep flattening the model tier while electricity prices and interconnection queues do not ease, profits will be recorded against energy and land. And if Chinese laboratories continue to open-source after reaching the frontier, openness is a long-term strategy; if they turn to closed commercialisation, it was a transitional instrument.

Three judgements constitute the main conclusions. First, the impact of artificial intelligence on distributive order begins with the failure of screening, and the reorganisation of the labour market follows; this ordering means that policy attention should include the reconstruction of measurement and certification, not only employment protection. Second, the boundaries of substitution are ultimately set by material conditions rather than algorithmic capability; a marginal cost of cognition approaching zero does not dissolve the price floor, so rent retreats to energy, land and institutionally created exclusivity, which constitutes the principal limit on expectations of general abundance. Third, distributive outcomes are determined by political process: rising productivity only extends the boundary of the possible and does not select where society lands. Historical precedents show that societies have rarely handled surplus populations decently, and that the least bad outcome is a coerced decency which has never been a gift from technology. As to whether institutions will advance, the honest answer is that they will not advance automatically, since institutional security cannot be produced by technical processes but only established by political ones.

Disclaimer:
This post was polished with the help of an LLM; wording may introduce ambiguity, and the author’s actual intent should prevail.

This English version was translated entirely by an LLM. Where any ambiguity arises, the Chinese original and its statement above take precedence.