How This Was Grown
This essay is not a clean handoff where I gave a prompt to a model and posted the result. It is an artifact from a live writing loop — less "prompt in, essay out" than growing an idea at a computer. One long GPT session held the evolving argument. Claude and a GPT Pro session read drafts as outside critics, and their useful objections went back into the main thread. A NotebookLM audio pass — two AI voices debating the draft out loud for twenty minutes — made it easier to hear where the logic was vague or a claim felt too confident. I chose the question, judged the feedback, and decided what belonged.
My AI work keeps running into the same question: if models become collaborators, tools, workers, and media engines, what happens to the value of human creative and cognitive labor? This essay came out of that loop, with me trying to keep the center of the idea intact.
Abstract
This essay argues that the central labor-market risk of advanced artificial intelligence is not unemployment alone, but the erosion of cognitive labor scarcity: the scarcity premium attached to economically deployable human cognitive work. The likely intermediate outcome is compression: fewer workers needed per unit of output, narrower career ladders, and a widening divide between those who sell ordinary cognition and those attached to ownership, authority, distribution, capital, and institutional trust. The decisive threshold is organization-level agency — the point at which one human can safely supervise many machine workflows — and the transmission mechanism is capitalist competition. The political danger is that routine cognition becomes cheap while the scarce complements around useful cognition remain concentrated: compute, energy, data, distribution, legal permission, housing, healthcare. A serious democratic response must own the bottlenecks, open the bottlenecks, guarantee agentic access, and build institutions of recognition. The argument is conditional rather than prophetic.
Prefatory Note: June 2026 and Scope
This essay is written from June 2026. That date matters. The argument is not made from a timeless view of "AI in general," but from a moment when frontier systems are visibly moving from chat toward agents, from task assistance toward workflow execution, from coding help toward software work, from model releases toward industrial-scale compute buildout, and from consumer products toward strategic infrastructure. The particular names will change. GPT-5.5, Codex, Claude Code, Fable, Mythos, AI 2027, and Stargate matter less as permanent reference points than as markers of a frontier state: by mid-2026, advanced AI is no longer only a text generator. It is becoming agentic, infrastructural, and geopolitical.
The essay's primary scope is affluent democratic capitalism, especially the United States, where employment is tightly linked to income, status, and political power. The global version of the argument — shortened development ladders abroad, new dependencies between nations — is taken up in Section VI.
One bracket should be explicit. This essay does not argue the alignment question — whether advanced systems can be kept under human control at all. It analyzes the branch in which control succeeds. That is a focus, not an evasion: the most-read AI-risk scenario of the moment, Daniel Kokotajlo's AI 2027, forks into two endings, and in the ending where the control problem is solved, the question that remains is exactly this essay's — who owns the systems, who governs them, whose preferences the abundance serves. If control fails, the labor question is moot. If control succeeds, the labor question is the whole question. The best case is a fight over ownership, and what follows is a map of it.
I. The Core Problem: Employment Is Not Leverage
Every major technological revolution produces a fear that machines will make human labor obsolete. Usually, that fear proves both understandable and incomplete. Machines destroy forms of work. Tractors reduced the need for farm labor. Assembly lines changed manufacturing. Computers displaced clerical calculation. Software automated bookkeeping, scheduling, logistics, and communication. Yet work did not disappear. It moved. Human beings adapted by shifting into new tasks, industries, and institutions. The standard reassurance about artificial intelligence begins from this history. Technology automates tasks, not work itself. Productivity rises, goods become cheaper, demand expands, and new jobs appear. Workers displaced from one domain move into another. The economy reorganizes, often painfully, but not catastrophically.
That reassurance is not wrong. It is incomplete.
Earlier automation did not leave cognition untouched. Mechanized production displaced skilled artisans. Computers automated calculation, clerical judgment, scheduling, routing, and information processing. Software absorbed many white-collar routines. Historical adaptation also depended on embodied service work, care work, mass consumption, and labor organization. The human fallback was not cognition alone. But broad human adaptability was one crucial fallback. When old tasks were automated, humans could often move toward other tasks requiring judgment, language, social intelligence, physical presence, institutional trust, or learning in new domains. Advanced AI is different if it begins to automate a large part of that adaptive layer: not only particular skills, but the ability to learn, coordinate, and execute new cognitive work across domains.
OpenAI's Charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work." That definition is useful not because it resolves the metaphysics of intelligence, but because it frames AGI as an economic and institutional threshold: autonomy plus economically valuable work.
The key concept is cognitive labor scarcity. By this I do not mean human intelligence in some abstract or spiritual sense. I mean the scarcity premium attached to economically deployable human cognitive task bundles: task execution, task learning, contextual judgment, verification, coordination, trusted interaction, accountability, and institutional authorization.
AI will not cheapen all of these equally. It may cheapen routine task execution before it cheapens judgment. It may cheapen first-pass drafting before it cheapens accountability. It may cheapen code generation before it cheapens software architecture, security responsibility, or product judgment. It may cheapen legal document preparation before it cheapens court appearance, client trust, or professional liability. The frontier is jagged.
But jaggedness is not a fixed map of human refuge. It is a moving frontier. Domains that are hard for one generation of models may become exposed in the next when models are combined with tools, memory, retrieval, verification systems, domain-specific data, multimodal perception, organizational redesign, and legal accountability wrappers. The long-term question is not whether AI becomes equally good at everything. It is whether even its weaker domains become economically sufficient.
If enough cognitive task bundles become cheaper to produce by machine, the bargaining position of many workers changes. The relevant comparison is not "human versus magical superintelligence." It is a firm asking whether to hire another junior analyst, paralegal, coder, or customer-support worker when an increasingly capable stack of models, tools, agents, and verification systems can perform much of the work.
AI is not the worker's outside option. More precisely, it is the employer's substitution alternative. As that alternative improves, it weakens the worker's bargaining position and may indirectly worsen the worker's own outside employment options. The occupation may still exist. The worker may still be employed. But wages, training pathways, status, and political leverage may decline.
The mechanism is a price ceiling, not a negotiation. When a machine can perform a task acceptably at a given cost, that cost caps what any human can charge for the same task, whatever the bargaining institutions around it. As the set of machine-performable tasks expands and machine costs fall, the ceiling covers more of the wage distribution, and human pay comes to rest on the shrinking bundle of tasks machines cannot do, are not trusted to do, or are not permitted to do. Productivity gains from working with machines raise wages only while the worker is needed to realize them. Bargaining language describes the transition. The ceiling is the mechanism.
A society can maintain high employment while the labor-based social contract weakens. People can still have jobs while those jobs no longer provide enough income, security, status, bargaining power, or dignity to support a broad middle class. The danger is not only that humans have no tasks left to perform. It is that human labor remains present but less economically necessary.
One clarification before the argument builds, because it changes what kind of claim this is. The claim is about leverage and distribution, not material decline. If machine cognition gets cheap, real consumption may well rise — cheaper software, tutoring, legal help, administration — even as bargaining power falls. That is not a comfort. It is what makes the problem politically hard: "you are materially fine" is exactly the argument that will be deployed against every redistributive claim in this essay, made to people whose leverage, status, and voice are eroding while their consumption basket improves. A society can become richer in goods and poorer in democratic labor power. The rest of the essay is about that second axis.
II. Organization-Level Agency
The labor-market threshold is not artificial consciousness. It is not even a model that can answer difficult questions. A chatbot answers. An agent acts. A machine organization coordinates.
Organization-level agency means machine systems can perform many of the coordination functions that make firms economically powerful. A model that writes a good memo is useful. A system that plans the work, gathers information, calls tools, coordinates subagents, drafts the memo, checks it, revises it, sends it, monitors the response, and updates the next step is closer to a worker. A system that coordinates many such workflows begins to resemble an operational department.
This threshold needs criteria. It is crossed not when a model sounds smart, but when machine systems can reliably:
-
maintain task horizons beyond single prompts;
-
coordinate heterogeneous tools, files, APIs, and agents;
-
recover from errors and exceptions;
-
verify intermediate outputs;
-
operate under changing or underspecified conditions;
-
reduce the frequency and cost of human intervention;
-
preserve audit trails and accountability handoffs;
-
communicate with institutions through authorized channels;
-
allow one human to safely supervise many machine workflows.
The supervisory ratio matters. If one human must carefully review every step, AI remains a productivity tool. If one human can safely oversee dozens or hundreds of workflows, AI becomes a labor-compressing organizational layer. The difference between augmentation and substitution often turns on supervision, verification, and exception handling.
The Stretch-Out
In 1800, a weaver ran one loom. By the 1830s, two. In 1842, Lowell pushed it to three — the workers called it "the stretch-out" — and by 1902 a single weaver in a Massachusetts mill tended eighteen. The economic historian James Bessen showed that the binding constraint, for the whole century, was never the machines. It was how many looms one human could watch.
In February 2026, an Anthropic engineer wired up sixteen Claude agents in parallel and mostly walked away while they built a working C compiler. Eighteen looms, sixteen agents: nearly the same number, a hundred and twenty-four years apart. The difference is the clock. The loom ratio took a century to move from one to eighteen. The measured autonomy of AI agents — how long they run acceptably without a human looking — is doubling roughly every four months on the latest estimates, down from seven across 2019-2025.
The Stretch-Out
machines per human overseer · log scale
Here is the strange part: nobody is measuring the ratio.
Not the labor economists — no official statistic anywhere tracks how many agents a human supervises, or what share of work has become reviewing rather than doing. Not the AI labs — Anthropic publishes rich telemetry on how often humans interrupt their agents and explicitly declines to prescribe ratios; Waymo disclosed that roughly seventy remote operators oversee three thousand robotaxis (one human per forty-three vehicles) and simultaneously withheld how often those vehicles phone home, calling the data "confidential business information." Not the academics — the formula for exactly this quantity has existed since 2004, in a corner of human-robot interaction research called fan-out: how many machines one person can sustain, computed from how long a machine runs acceptably unattended versus how long each human touch takes. Twenty-two years of validation on robots. It has scarcely been applied to language-model agents.
So let me sharpen the definition this essay has been using. Define supervisory intensity — call it σ — as the human hours spent on a fleet of machine workflows (specifying, monitoring, reviewing, cleaning up) divided by the fleet's hours of operation. Its inverse is the supervisory ratio itself: how many agent-hours one hour of human attention sustains. The crucial design choice is that σ can be computed from numbers companies already leak. Firms guard the ratio like a trade secret, but they publish rates constantly — escalation rates, interrupt rates, review-sampling rates, the percent of AI pull requests a human merges. Multiply each rate by its handling time — the one number you must estimate — sum, divide by the fleet's operating hours, and you have σ. The metric the industry won't disclose is derivable, to a defensible estimate, from the metrics it brags about.
The early readings point in both directions — the jaggedness thesis in miniature. In June 2026, Anthropic disclosed that Claude writes more than eighty percent of the code merged at Anthropic. That is not a ratio — it is a share of output, silent on how much human steering each merged line consumed — but it is the strongest available proxy, from the firm best positioned to run the experiment: engineers there no longer primarily write code; they select, steer, and verify it. Pointing the other way, the METR study discussed in Section IV is also a ratio measurement — experienced developers on complex open-source tasks were effectively operating below one to one, spending more time supervising than the machine contribution returned. Both findings are real. The ratio is high where verification is cheap and integration is owned, and below break-even where context is deep and errors are costly. The question is which regime spreads. Capability tells you what machines can do; the ratio tells you what firms no longer need people for.
There is a second reason the labs are where the ratio moves first, and it changes how this whole dashboard should be read. Automating their own research is the frontier firms' stated strategy — legible in their disclosures, and reported from the inside by departed researchers, most publicly Kokotajlo — driven less by margin competition than by a race among a handful of firms to reach the frontier before each other. The sequence those firms describe is not diffusion outward and then improvement inward; it is the reverse: automate the production of machine cognition first, take the resulting systems to the wider economy after. If that sequencing holds, lab-internal ratios are not merely the best-instrumented corner of the economy. They are the leading edge of the entire curve, and every economy-wide indicator in this essay lags them by design. Section IV returns to what that possibility does to the scorecard.
Other proxies are available now to anyone who looks: headcount per shipped product in software firms year over year; the ratio of agent seats to human seats in enterprise AI contracts; the share of output in a workflow that receives step-level human review versus outcome-level sampling; job postings for "agent operations" and exception-handling roles, which are the residue the ratio leaves behind as it rises. None of these is clean. Together they are a dashboard.
Why does one number matter this much? Because the two futures this essay has been weighing make opposite predictions about it. If cognition is genuinely becoming cheap — if the compression story is right — then σ falls, sector by sector, year over year: each human supervises ever more agent-hours, and the human share of cognitive output collapses toward the supervisory residue. But the strongest counterargument to this essay says oversight is where the collapse stops: that reviewing, verifying, and owning the liability for machine output is the scarce complement that cannot be automated — in which case σ floors, and the wage story becomes a supervisor premium rather than a compression. Falling σ or floored σ. One number, tracked honestly and read alongside the postings and wage indicators in the Section IV scorecard — falling σ alone proves productivity, not compression — adjudicates between me and my best critics. I intend to score it either way by January 2030.
The precedents tell us how to track it honestly. Nuclear regulators codified operators-per-reactor as a literal staffing table, and when NuScale won approval to run twelve reactor modules with six operators in 2023, the compression event had a name, a date, and a simulator-validated evidence file. Aviation collapsed the five-person cockpit to two over three decades — radio operator, then navigator, then flight engineer — with the last step settled through defined workload factors and one blue-ribbon adjudication in 1981. Supervisory ratios in serious industries move as step functions anchored by named decisions — and the AI equivalents are already being drafted unnamed: the FAA's proposed Part 108 would put a single "flight coordinator" over drone fleets — its delivery permits run to a hundred aircraft — with the actual ratio left to manufacturers' operating manuals. Anyone building the σ time-series should log these events the way the nuclear industry logs exemptions, because in twenty years they will be the chapter markers of the stretch-out.
Four honesty clauses, so the metric can't be gamed into optimism. First, σ is meaningless without a quality index: a falling ratio with a rising defect-escape rate isn't automation, it's deferred rework — the loom era knew this too, which is why Bessen found the mills tripling training investment per worker as the ratio rose. Second, declare the boundary: the Air Force's Reaper drone needs two people in the control station and roughly fifty across the whole enterprise per airframe. Count only the cockpit and every ratio flatters itself by an order of magnitude. Third, the denominator counts only agent-hours that produced output someone accepted: park a fleet in idle loops overnight and σ falls while nothing real was automated. Fourth, an agent-hour is not a constant unit of work the way a loom-hour was a constant length of cloth: a model that produces the same accepted output ten times faster shrinks the denominator tenfold and makes σ read ten times worse, so σ must be scored alongside output per agent-hour — otherwise every speedup counts as a step backward.
σ — Supervisory Intensity
human hours per fleet-hour of machine work · derived from the rates firms already publish
The weaver watching eighteen looms in 1902 was the most productive textile worker in human history. Within a generation the mills went south and the towns built on them emptied. That is the whole thesis of this essay in one image: productivity is not protection. The stretch-out is beginning again, at four-month doubling time, and this time we have the instruments to watch it happen — if anyone bothers to point them.
Measurement is one half of the threshold. Accountability is the other. A firm is not just a coordination graph. It is a legal and institutional entity. It can own assets, sign contracts, hire workers, insure risks, sue and be sued, pay taxes, hold licenses, and bear responsibility. If accountability remains a durable human or corporate bottleneck, then machine systems may not become firms in the legal sense even if they become organization-like in the operational sense.
The likely near-term structure is therefore not fully autonomous AI firms floating outside law. It is operational machine organization inside legal wrappers, with responsibility assigned to corporations, licensed professionals, executives, insurers, auditors, or state-certified operators.
AI may become operationally organizational before it becomes legally organizational.
This reframes a central tension. Human beings may remain valuable as accountability nodes even as machines absorb much of the coordination work. But that does not mean human labor keeps its old leverage. The high-trust human layer may shrink into sign-off, governance, relationship management, liability, exception handling, and institutional representation. The work remains human-adjacent, but the old staffing pyramid compresses.
Over time, even accountability may be partly reassigned. Insurance pools, audit systems, regulatory certifications, corporate liability regimes, and professional standards may adapt to machine workflows. The question is not whether machines become legal persons. The question is how much human supervision law and legitimacy require per unit of machine output. If that ratio falls, labor leverage falls with it.
III. From Capability to Political Economy
The path from AI capability to labor devaluation is not automatic. It runs through a causal chain: capability -> cost and reliability -> institutional deployment -> substitution or augmentation -> organizational redesign -> hiring and wage effects -> demand response -> labor's share of income -> political bargaining power.
Every link is contingent. A system may be capable but too expensive. It may be cheap but unreliable. It may be reliable but legally constrained. It may automate some tasks while complementing others. It may lower output costs but expand demand enough to preserve employment. It may raise productivity while concentrating the gains. The goal is not to predict one future, but to identify the conditions under which different futures become likely.
Capitalist Competition as the Transmission Mechanism
AI capability does not reorganize labor by itself. It becomes labor pressure through competition.
In a capitalist market, a firm that can produce acceptable output with fewer workers, lower costs, faster iteration, or greater scale gains an advantage. If AI allows one firm to compress a workflow while maintaining quality, rivals face pressure to imitate it. A company may preserve a larger human staff for good reasons — trust, training, culture, quality, liability, or brand — but it carries a cost if competitors achieve similar output with smaller AI-augmented teams. Over time, what begins as optional efficiency can become an industry standard.
The key variable is not model capability alone. It is total cost of usable output. AI substitutes for labor when generation, supervision, verification, integration, liability, management, and customer-trust costs together fall below the cost of the human workflow. In some domains, AI will look cheap but remain expensive because errors are costly or review is hard. In others, AI will become overwhelmingly cheaper because verification is easy, automated, or shifted to a smaller expert layer.
This means jaggedness is partly a total-cost phenomenon. A task becomes exposed when AI capability plus verification plus integration becomes cheaper than the human workflow. The frontier moves not only when models improve, but when firms restructure work to make outputs easier for machines to generate, check, and deploy.
The competitive mechanism often works as a ratchet. First, some firms experiment. Then a few discover lower-cost workflows. Investors, executives, and competitors update their expectations. Firms that do not adopt begin to look bloated and slow. Hiring targets shift. Junior roles are not backfilled. Departments flatten. Vendors advertise AI-native cost structures. What began as "AI adoption" becomes the ordinary way to run the business.
This pressure differs by market structure. In competitive markets, firms may pass savings to consumers through lower prices. In oligopolistic markets, firms may keep prices high and capture the surplus. In regulated sectors, liability, licensure, public trust, and compliance slow adoption. In public sectors, budget pressure and procurement rules matter more than profit competition. In luxury or status markets, human labor may remain valuable because the human element is part of the product.
The mechanism is therefore not that capitalism automatically replaces every human wherever possible. Capitalist competition creates a persistent bias toward labor-saving adoption when machine-mediated production lowers total cost without unacceptable losses in quality, legality, trust, or control.
This also changes bargaining before full replacement occurs. Management gains a credible threat: more of this work can be automated. That threat can discipline wage demands, weaken worker confidence, and shift internal politics toward executives and owners even when substitution remains partial.
Three scenarios matter most: augmentation, compression, and post-labor substitution.
- Augmentation: supervisory burden remains high, humans stay central, verification requires expertise, demand expands enough to preserve hiring, AI is layered onto existing jobs, and gains may be shared.
- Compression: supervisory burden falls, fewer humans oversee more output, verification becomes increasingly automatable, demand expands without preserving staffing ratios, organizations redesign around machine workflows, and gains concentrate among firms and top workers.
- Post-labor substitution: many workflows run with limited human oversight, verification becomes low-cost or machine-mediated, output expands mainly through machine production, staffing pyramids shrink, and productive ownership dominates wage income.
These scenarios can coexist. AI may augment elite professionals, compress routine cognitive occupations, substitute for some services entirely, and increase demand for embodied care at the same time. The future will not be one clean regime. It will be sectoral, jagged, and politically mediated.
Pace matters. If the transition unfolds over fifty years, demographic turnover, new training pathways, institution-building, and political adaptation have more room to work. If it unfolds over ten or fifteen years, the broken-ladder and bargaining-power effects are sharper. The strongest version of the labor-compression argument applies under rapid capability improvement, falling inference costs, low supervisory burden, fast organizational redesign, and concentrated ownership of the complementary bottlenecks. Machine cognition does not need to become free to devalue labor. It only needs to become good enough, fast enough, reliable enough, and cheap enough to change the employer's substitution alternative.
The Counterfactual: What Machine Cognition Adds
A fair objection: labor's position was deteriorating long before transformers. Labor's share of income has declined since roughly 1980, driven by deunionization, globalization, market concentration, and policy — four decades of weakening with no AI required. Offshoring, in particular, was already the employer's substitution alternative for an earlier generation of cognitive work. So what does machine cognition add that the twentieth century had not already accomplished? Four things. Speed: an offshore operation takes years to stand up; an agent deployment takes a quarter. Reach: offshoring touched tradable services; machine cognition reaches the non-tradable core — the local law office, the regional hospital's back office, the small firm that could never afford a Bangalore contract. Friction: the offshore alternative carried coordination costs, time zones, quality management, and political exposure; the machine alternative carries an API bill. And trajectory: wage arbitrage was a one-time level shift that narrowed as poor countries grew richer, while machine capability compounds on an improvement curve with no equivalent equilibrating force. AI does not begin the erosion of labor's position. It removes the frictions that slowed it and the equilibration that eventually checked it. That is why this substitution alternative deserves its own analysis rather than a footnote to the last one.
IV. Evidence, Rising Jaggedness, and a Software Case Study
The empirical record does not prove a post-labor future. It supports plausibility, not certainty. Current evidence shows task-level productivity effects, uneven adoption, exposed junior work, and early organizational experimentation. It does not yet establish broad wage compression, a declining labor share caused by generative AI, destroyed career ladders at scale, or durable class bifurcation.
The best reading of current evidence is jagged, but jaggedness should be understood dynamically.
Generative AI can produce meaningful gains in some cognitive tasks, especially where work is digital, repeatable, tool-mediated, and cheaply verifiable. It can also mislead workers, degrade judgment, increase review costs, or slow experts down when tasks are complex, poorly specified, or outside the system's competence.
This unevenness is real. But as Section I argued, jaggedness determines the sequence of exposure more than the final boundary of human advantage.
The strongest empirical anchors point in different directions, which is exactly why the jaggedness frame matters.
In a large study of 5,172 customer-support agents, Erik Brynjolfsson, Danielle Li, and Lindsey Raymond found that access to a generative AI assistant raised productivity by roughly 14-15 percent on average, with the largest gains for less experienced and lower-skilled workers. That supports the skill-compression thesis: AI can transfer patterns from stronger workers to weaker workers and flatten some experience premiums. A Harvard Business School / Boston Consulting Group field experiment found a different but compatible pattern. Consultants using GPT-4 completed 12.2 percent more tasks, worked 25.1 percent faster, and produced work rated more than 40 percent higher in quality on tasks within the AI frontier. But performance worsened on tasks outside the frontier. This supports the idea that expertise still matters when the key skill is knowing when not to trust the tool.
Coding evidence is also mixed. In a controlled GitHub Copilot experiment, developers using Copilot completed a JavaScript HTTP-server task 55.8 percent faster than the control group. But a METR randomized study of experienced open-source developers in early 2025 found the opposite in a more complex setting: AI tool use made tasks take 19 percent longer, despite developers believing beforehand that AI would speed them up. That result does not refute the labor-compression thesis. It sharpens it. AI is labor-saving only when the cost of supervision and verification is low enough.
The evidence therefore points to a more precise claim: AI compresses skill where tasks are inside the frontier and verification is easy. It amplifies expertise where knowing the frontier matters. It disappoints where output is hard to verify, context is deep, and errors are costly. But those categories are not permanent. Tool use, memory, retrieval, synthetic data, self-checking, formal verification, multimodal perception, and better institutional integration can move tasks from "outside the frontier" to "inside the frontier."
There is also a macroeconomic version of the skeptical case, and it deserves stating plainly. Daron Acemoglu and others have argued that the measurable aggregate effects of generative AI remain small: modest productivity gains, a labor share that has not visibly moved because of it, and adoption that is wide but shallow. On this view, the compression thesis extrapolates a software-sector anomaly to the whole economy. The honest response is that the skeptics are right about the present. Nothing in current national statistics shows compression at scale. The disagreement is about mechanism and horizon: whether task-level exposure plus competitive pressure plus organizational redesign eventually shows up in the aggregates, or dissipates the way many general-purpose-technology forecasts have. That is exactly why this essay stakes itself to dated observables below, rather than to the aggregate record so far.
Software engineering is the clearest sectoral case because the causal chain is visible.
At the capability layer, coding models can now produce functions, tests, refactors, documentation, and first-pass architectural suggestions. Code has an advantage over many forms of knowledge work: it can often be run, tested, linted, benchmarked, reviewed, and deployed through existing tooling. This makes it one of the first domains where AI can move from suggestion toward execution.
At the cost and reliability layer, the important question is not whether the model can write code. It is whether generated code reduces total project cost after review, integration, debugging, security, maintenance, and coordination. A fast code generator that creates subtle bugs may increase total cost. A slower but more reliable agent that passes tests, explains changes, and integrates with the repository may reduce it. The supervision burden is decisive.
At the deployment layer, firms first layer AI onto existing developers. Engineers use copilots, chat interfaces, coding agents, documentation tools, and test generators. This is augmentation. The worker remains the central producer and judge.
At the organizational-redesign layer, the question changes. If senior engineers can use agents to perform tasks once assigned to juniors, firms may reduce entry-level hiring or expect smaller teams to ship the same product. A startup that once needed ten engineers may try to operate with three senior engineers and a large tool stack. A large firm may keep platform, security, product, and architecture teams while thinning routine implementation roles. The profession survives, but the ladder narrows.
At the demand layer, cheaper software can expand output. More internal tools, prototypes, niche apps, automations, and small products become viable. AI may enable one-person software firms and small teams that could not previously build at professional scale. This complicates the pessimistic story. The same technology that compresses employment inside incumbent firms may broaden entrepreneurship outside them.
At the ownership layer, the gains depend on who controls distribution, cloud infrastructure, model access, and user trust. If AI makes code cheap but distribution remains scarce, then value migrates from coding labor toward platforms, product ownership, customer access, and infrastructure. If open tools, cheap hosting, and interoperable distribution improve, some value may diffuse to small firms and individuals.
Software therefore illustrates the entire argument. AI may augment developers, compress teams, break or redesign apprenticeships, stimulate new demand, broaden micro-enterprise, and shift rents toward infrastructure and distribution. The net effect cannot be inferred from model capability alone. It depends on verification cost, supervision ratios, demand elasticity, firm redesign, and ownership of complements. It also illustrates why jaggedness is not a permanent refuge. Software is exposed early because verification is unusually available. Other fields may follow more slowly, but not necessarily never. As medicine, law, finance, education, design, science, and administration become more tool-mediated, data-rich, and auditable, the same dynamics can spread. The sequence differs by domain. The underlying mechanism is broader.
The current evidence therefore supports urgency under uncertainty.
What Would Change My Mind
An argument this conditional owes its readers accountability. Conditions can become a hiding place: if every outcome is compatible with the framework, the framework predicts nothing. So the compression thesis should be staked to observables and dated.
I will treat the thesis as failing — wrong, not early — if by January 2030 the following picture holds. Entry-level postings in AI-exposed cognitive occupations recover to within 15 percent of matched non-exposed occupations from the same 2022 baseline while overall employment stays healthy — prime-age employment-to-population at or above 80 percent. The wage premium for mid-skill cognitive work over embodied service work stabilizes or widens. And supervision stays stubborn: outside demonstrations and marketing, no major cognitive sector operates routine production at supervisory ratios above ten machine workflows per human overseer. That world would mean verification costs are the durable wall this essay feared they would not be. The correct conclusion in that world is not that compression is delayed. It is that augmentation won, that human judgment priced itself back into the loop, and that the policy urgency of Sections VIII through X was overdrawn.
I will treat the thesis as confirmed in its strong form if the opposite picture holds. Entry-level postings in exposed occupations remain 30 percent or more below matched occupations while output in those sectors grows. The cognitive wage premium continues to compress. At least two major cognitive sectors run routine production at supervisory ratios above ten to one. And firm formation shows the signature shape: rising output per firm with falling headcount per firm in cognitive industries.
There is a third picture, and it needs pre-registering more than either pole, because without a name it would read as thesis-failure while being closer to the opposite. Call it the takeoff route. Competitive diffusion, sector by sector, is not the only way cognitive scarcity can end. The labs' stated sequencing — automate AI research first, diffuse afterward — is the other, and it is the shape the scenario literature predicts, most concretely Kokotajlo's AI 2027: broad indicators that stay quiet — postings recovering, the wage premium holding, no ordinary cognitive sector near ten to one — while the anomaly concentrates in the one sector that produces machine cognition. That anomaly has a checkable signature: revenue and shipped output at frontier labs decoupling from lab headcount; lab-internal supervisory ratios collapsing while economy-wide ratios sit still; the verification layer inside the labs itself automating, machine work graded by machines. The scorecard carries this as a sixth indicator, added in July 2026 — a month after publication, which is why it is dated here; the original five indicators and both poles stand unmoved. If January 2030 shows quiet broad indicators and that lab anomaly burning, the honest reading is not that compression failed. It is that compression was mispredicted in shape — concentrated in the toll-road sector rather than spread across the economy — with the diffusion wave queued behind it and the political window of Sections IX and X mostly closed. But the route has to actually fire to be claimed. If the broad indicators are quiet and the lab anomaly is quiet too — lab headcount and output rising together, lab ratios stubborn — then augmentation won everywhere, including where the strategy said it would win first, and the thesis is wrong, not early, with no third door to hide behind.
Five measurement rules keep the scorecard honest. The 2022 baseline is a bubble year for cognitive hiring, so postings are read as exposed occupations against matched non-exposed occupations from the same baseline — the thesis needs exposed work to fall relative to its peers, not merely relative to a peak. Which occupations count as exposed is pre-registered, not chosen at scoring time: the scorecard uses a published exposure index — Eloundou et al. as primary, Felten et al. as the robustness check — and names its postings sources in advance. The wage premium is read from new-hire wages at fixed percentiles, because compression that destroys the bottom of cognitive work would raise the measured premium of the survivors and read as thesis-failure while the thesis was coming true. The supervisory-ratio tests use the σ method above, at the enterprise boundary, and count only sectors of cognitive production — the same exposure universe as the postings indicator — so the number means the same thing each time it is scored and a robotics milestone cannot confirm a thesis about cognition. And the scorecard owns its blind spot: none of the indicators above would see displaced workers absorbed into task categories that did not exist in 2022 — which is how every previous wave partially resolved — so it carries a fifth indicator, employment and posting share in occupation titles that did not exist at baseline. If the new-task channel runs strong, the compression thesis weakens no matter what happens to the old categories.
Between those poles lies the jagged middle this essay considers most likely — in which case the honest scorecard is indicator by indicator, updated in public, not a single verdict. I intend to keep that scorecard, with one aggregate rule so that neither the jagged middle nor the takeoff route can become a hiding place: if a majority of the five labor-market indicators land on the augmentation side by January 2030 and the sixth — the takeoff signature — has not fired, I will call the thesis wrong, not early. The point of writing from June 2026 is that June 2026 can be checked.
The scorecard is live: Compression Scorecard.
V. Comparative Advantage, Demand, and the Ladder
The strongest economic objection to labor pessimism is comparative advantage. Even if AI becomes better than humans at most cognitive tasks, humans do not automatically become unemployable. A less productive agent can still trade with a more productive one if opportunity costs differ. Humans may remain useful where machine intelligence is constrained by compute, law, trust, physical-world execution, personal presence, or demand for human interaction.
This objection is important. It prevents sloppy claims that human labor must vanish the moment machine labor becomes superior. But comparative advantage can preserve employment without preserving the wage premium of human cognition.
The wage a worker commands depends on marginal productivity, bargaining power, and alternatives. If machines can perform many cognitive tasks cheaply, the market wage for human cognitive labor is disciplined by the employer's ability to substitute. Humans may still be employed doing what machines cannot do, what institutions require humans to sign for, what buyers prefer humans to perform, or what is not worth automating fully. But those roles may command less pay, status, and leverage than the old knowledge-work middle class.
Forecasters who take machine capability furthest arrive at the same place from the other direction. Asked which jobs survive full automation, Kokotajlo answers that it is a political question, not a technical one: at sufficient capability, what remains human is what law, licensure, and buyer preference reserve — the judge whom statute requires to be human, the nanny whom parents simply prefer human. The list of protected roles is written by institutions, not by the frontier.
Demand expansion complicates the picture. If software becomes cheaper, society may want more software. If tutoring becomes cheaper, students may receive more instruction. If legal help becomes cheaper, more people may pursue legal remedies. Lower prices can create new markets.
But demand for cognitive output is not the same as demand for human cognitive labor. Output can expand while human labor share falls if AI satisfies much of the new demand. A world can have far more software, tutoring, legal drafting, entertainment, analysis, and research while needing fewer human workers per unit of output.
Personal agents also cut both ways. A household's agent may bypass paid intermediaries by drafting letters, comparing quotes, filling forms, or preparing legal claims. That can remove paid service transactions from the demand side. But the same agent may also stimulate demand. A health agent may generate more doctor visits by noticing problems earlier. A legal agent may uncover valid claims that would otherwise go unpursued. A home-maintenance agent may create more repair work by detecting failures before they become disasters. The net effect depends on whether the agent substitutes for the professional's core contribution or primarily reduces search, coordination, and administrative friction.
The apprenticeship ladder is similarly ambiguous. AI may break the ladder if firms use it to avoid hiring juniors — Section VI takes up that mechanism. But AI may also repair or redesign the ladder. It can simulate cases, tutor novices, provide feedback, accelerate practice, and transfer expert patterns to less experienced workers. Professional organizations may deliberately preserve junior roles because they need future seniors. The broken-ladder argument is strongest when firms capture short-term savings by reducing junior hiring faster than institutions create new training pathways.
There is also the possibility of AI-enabled micro-enterprise. If AI lowers the cost of creating software, media, research, education, marketing, and services, individuals and small teams may become more productive. One-person companies and small cooperatives could proliferate. This would not preserve traditional wages, but it could spread productive agency.
The constraint is that micro-enterprise still needs complements: distribution, trust, payment systems, legal permission, customer acquisition, reputation, capital, data access, and attention. AI can make creation easier while leaving the gates to market power intact. The question becomes who controls the gates between making something and it mattering.
Finally, lower wages do not automatically mean lower material welfare. Cheaper services can raise real consumption even as leverage falls — the divergence Section I put at the center of this essay. Again: a society can become richer in goods and poorer in democratic labor power.
Cost Disease, Deployed
There is one mechanism working quietly in labor's favor, and it deserves naming: cost disease. As Baumol observed, when productivity surges in one part of the economy, the sectors that cannot be automated get relatively more expensive — the string quartet still takes four musicians. If machine cognition floods the zone, the human-remaining sectors — hands-on care, embodied trades, in-person teaching, live performance, anything where the human presence is the product — should see their relative prices rise. That is a real channel through which some human work gains value in an AI economy, and the augmentation scenario leans on it. But cost disease cuts twice. The same dynamic that raises the price of human-delivered care also raises the cost of everything the AI dividend is supposed to buy. Housing, healthcare, education, and eldercare are precisely the Baumol sectors — and precisely the rent-extracting sectors of Section VIII. A society can watch machine cognition make information services nearly free while the human-bottlenecked necessities absorb every gain.
And a rising price is not a rising wage. Displaced cognitive workers have to go somewhere, and where they go is into these same human-remaining sectors — a labor-supply shock aimed at the last places machines cannot reach. Baumol raises what the service costs. Whether the worker delivering it captures the difference depends on entry. Where entry is gated — a license, capital, a client list — incumbents harvest the premium. Where entry is open, the crowd competes the wage back down even as the price rises, and the wedge goes to whoever owns the gate. The barber and the nurse captured a century of manufacturing productivity because workers reallocated slowly into their trades. A fast, wide cognitive displacement runs the same channel in reverse: the queue outside the human sectors holds the wage down while the price of being served keeps climbing. Whether cost disease becomes labor's refuge or the landlord's harvest depends, once again, on who owns the bottleneck.
VI. Labor Compression and Class Bifurcation
AI can compress labor through several mechanisms at once: broken or redesigned ladders, smaller firms, temporary scaffolding, and agentic disintermediation.
Entry-level knowledge work is especially exposed because it is often composed of reviewable first-pass tasks. AI can draft, summarize, search, classify, generate options, and produce rough versions quickly. If firms substitute AI for junior work, they may save money now while undermining future talent formation. If they use AI as training infrastructure, the ladder may shorten rather than break. Which happens depends on incentives: short-term cost reduction versus long-term professional reproduction.
The broken ladder is also a collective-action problem. Industries need experienced workers eventually, but each firm has an incentive to reduce the entry-level work that trains them if competitors are doing the same. A firm may know that apprenticeship matters in the long run while still cutting junior roles in the short run because AI-native competitors operate with lower staffing ratios. Everyone needs seniors later; no individual firm wants to carry the cost of training them now.
Firms also redesign work around machines. They do not simply ask whether AI can perform the old human job exactly. They standardize inputs, constrain outputs to templates, build approval pipelines, automate verification, move judgment to a smaller expert layer, force customers into structured interfaces, create machine-readable policies, and shift liability to reviewers or users. The production process changes so machines can handle more of it. The relevant question is not "Can AI do this whole job?" but "Can the firm redesign the job so AI can do enough of it?"
Firms may also compress. A firm is an information-processing and coordination system: goals, task assignment, memory, communication, quality control, accountability, and adaptation. AI becomes economically radical when it targets not only task execution but coordination itself. Many firms may not disappear. They may become smaller and change shape. The old pyramid of white-collar labor becomes a narrower control tower: owners, trusted decision-makers, clients, and legal accountability at the top; fleets of machine agents in the middle; a thinner human layer around oversight, relationships, exceptions, and legitimacy.
It matters what those middle fleets are. Dario Amodei's phrase for the destination is a "country of geniuses in a datacenter"; Kokotajlo's correction is that it would be an army, not a country — copies of one model, centrally updated, owned by one firm, under one command structure. A workforce like that cannot organize, strike, quit, or slow-walk. When one copy learns a skill, every copy has it, so the training pipeline that made junior humans expensive and senior humans scarce is simply absent. The middle of the compressed firm is staffed by labor whose bargaining power is exactly zero, and every human negotiation in the building happens against that baseline.
AI will also create jobs: workflow designers, agent managers, model evaluators, and compliance reviewers. Some will endure because they attach to durable scarcities: accountability, governance, institutional judgment, trust, physical execution, or legal authority. Others may fade as systems become easier to use, more autonomous, and more deeply integrated.
The class result is not flat labor devaluation. It is skill compression inside some tasks and class bifurcation around scarce complements.
AI may reduce differences between novice and expert performance where tasks are standardized and feedback is dense. A novice support worker can sound more experienced. A junior programmer can complete routine tasks faster. A generalist can produce competent drafts in unfamiliar domains. This compresses some returns to ordinary expertise.
But skill compression can coexist with class divergence. The partner who owns the client relationship, the lawyer who appears in court, the doctor whose judgment carries liability, the executive who controls deployment, the creator with distribution, the firm with proprietary data, and the owner of compute infrastructure occupy different positions from workers producing routine cognitive output. AI can level some task performance while widening the gap between those who sell ordinary cognition and those who control ownership, authority, trust, legitimacy, and distribution.
Care work reveals the same principle. It is tempting to say care, education, and human presence will become more valuable because they remain human. That may be true in demand terms. But necessity does not automatically create power. Care work has always been necessary and often underpaid. The crowding mechanism from Section V makes it worse: every displaced cognitive worker who retrains toward care adds to the supply that holds its wage down. Wages rise only if demand is backed by purchasing power, public funding, labor organization, and social norms that convert necessity into bargaining power.
The future high-trust layer may therefore not be large enough to preserve a broad middle class. It may include licensed professionals, care workers, educators, human-facing service providers, managers, artists, performers, community leaders, and accountability holders. But whether these roles support mass dignity depends on institutions, not on human scarcity alone.
The class line runs between those who own or control the systems around useful intelligence and those who merely use them.
This Essay as Its Own Case Study
I should note that this essay is itself a data point. The process described at the top — one person steering a stack of models through drafting, critique, cross-examination, and audio review — produced in days what would recently have required weeks of a researcher's labor, or a researcher plus an editor plus a reading group. The writing labor was compressed. But look at where the human role migrated rather than vanished: to the choice of question, the judgment of which objections mattered, the taste that killed weak drafts, and the ownership of the byline and the site you are reading it on. Selection, steering, accountability, and distribution — the exact residue the compression thesis predicts. And the value produced here was captured by the owner of the taste and the domain name, not by the systems that generated most of the sentences. I am not exempt from my own argument. I am an early instance of it.
The Shortened Ladder Abroad
The global version of this argument may matter more than the domestic one, and it deserves more than a gesture. The late-twentieth-century development ladder ran substantially through exported cognition: call centers, business-process outsourcing, translation, medical transcription, back-office finance, and outsourced software. The Philippines built an industry of well over a million business-process jobs on it; India built a middle class on it. These were the rungs by which countries without manufacturing bases converted English fluency and education into foreign exchange and upward mobility. They are also, nearly rung for rung, the tasks most exposed to current machine cognition: digital, repeatable, tool-mediated, and cheaply verifiable — the exact profile of early exposure from Section IV. If those export markets thin, the ladder shortens for the countries that need it most, at the same time as frontier compute, chips, energy, and model access concentrate in a handful of firms and states. The domestic argument of this essay — cheap cognition, concentrated complements — reappears between nations with the stakes raised: countries, not workers, facing a world that needs less of what they sell while the new scarce inputs sit behind someone else's export controls.
A serious international agenda would extend the same four imperatives across borders: participation in ownership of frontier infrastructure, open access to capable models as development policy, guaranteed agentic access as a diplomatic instrument, and recognition that the next development ladder must be built deliberately, because the old one is being automated away from under the countries still climbing it.
VII. Life After Economic Necessity
Ten or twenty years after advanced AI begins reshaping cognitive labor, the strange fact may not be that humans have nothing to do. It may be that humans are intensely active while being less economically necessary. People may teach, perform, organize, build local projects, compete in games, worship, and argue about politics. The difference is that fewer of these activities may be reliably connected to wages.
Work is not only an income machine. It is a meaning machine. It gives people a role, a schedule, a standard of competence, a hierarchy of recognition, and a reason other people depend on them. This does not make work sacred. Much work is alienating, coercive, pointless, or humiliating. But modern society has used work as its default answer to the question of why an adult matters.
The post-labor problem is therefore an unbundling problem. Work bundled money, role, routine, status, competence, recognition, responsibility, and adulthood into one institution. A post-labor society has to rebundle those functions elsewhere. Cash alone does not provide a role. Tools alone do not provide recognition. Leisure alone does not provide responsibility. A society that solves income but fails to create recognized forms of contribution may produce comfort without dignity, freedom without direction, and activity without weight.
One possible future is agentic abundance. People receive income from public claims on AI capital, use machine intelligence as a general-purpose extension of agency, and work becomes less survival-driven and more about contribution, care, taste, curiosity, and belonging.
Another possible future is the busy-but-not-needed society. People are not lounging in leisure. They are managing agents, making content, doing gig work, caring for family, and trying to stay relevant. The post-labor world may not feel like idleness. It may feel like constant activity without clear social necessity.
A third possible future is rent-dependent agency, the world Section VIII maps. Everyone has AI, but not equal AI: the wealthy hold high-autonomy agents plugged into capital, legal support, and institutional permissions, while the poor get restricted, surveilled, ad-supported ones. Daily life gets easier in some ways, but dependency deepens.
The opposite of alienated labor is not passive leisure. It is free contribution under conditions of dignity. A humane post-labor society would therefore need institutions of recognition: places where people are needed, seen, challenged, trusted, and able to contribute. Care, education, art, civic service, and community would be part of the new social infrastructure of meaning, not extras bolted on after the economy is solved.
The question is whether the activities left to humans are connected to dignity, income, community, and power, or whether they become private hobbies inside a system owned by others.
VIII. Rent Dependency and Ownership Regimes
If routine cognition becomes cheap, intelligence itself may not be the main source of rents. A commodity whose marginal price falls cannot easily sustain a toll by itself. The rent migrates to the systems that make intelligence useful.
The toll road is not generic intelligence. The toll road is the scarce system around useful intelligence. Those systems include frontier compute, chips, energy, cooling, cloud platforms, proprietary data, model access, distribution channels, identity systems, payment rails, institutional APIs, legal permissions, housing, healthcare, land, and state-security infrastructure. Cheap intelligence becomes powerful only when it can act through these bottlenecks.
This is better described as rent dependency than as literal feudalism. "Technofeudalism" is a useful warning image, but the precise mechanism is simpler: every road into economic life can become a private toll road.
The rent argument must answer a hard question: why would these rents persist instead of being competed away?
Some layers may commoditize. Open models, model efficiency, edge inference, specialized chips, standards, competition among hyperscalers, and public regulation could erode model and inference rents. A world of decentralized, cheap, capable AI is possible. The rent-dependency thesis does not require permanent scarcity of every model layer.
It requires durable concentration in at least some complements. That is plausible where bottlenecks involve massive fixed costs, energy and grid access, proprietary data, vertical integration, and regulatory capture.
This is also where the essay's two halves have to be reconciled, because the argument needs cognition cheap and some of its inputs scarce at the same time. The reconciliation is a two-tier structure, and it already exists: cloud computing. Marginal compute is cheap enough to discipline anything that competes with it, while the capital stack above it extracts rent not through unit price but through access terms, priority, reserved capacity, and vertical integration. Machine cognition can work the same way: inference cheap enough to weaken wages, with the rent collected at the layer of who gets capacity first, on what terms, integrated with what. And the electricity analogy should be followed to its endpoint rather than abandoned halfway: kilowatt-hours became cheap partly because the natural-monopoly grid was rate-regulated with capped returns. That endpoint is not a rebuttal of the rent thesis. It is the historical precedent for Section X's second imperative — bottlenecks this essential end up either publicly constrained or privately harvested.
The Stargate buildout is a useful marker. The point is not that one project proves the future. It is that frontier AI firms are behaving less like ordinary software companies and more like builders of industrial-scale intelligence infrastructure. If the production of machine cognition depends on enormous capital expenditure, energy access, land, chips, cooling, and grid interconnection, then the ownership of those systems becomes central to political economy.
Housing and healthcare should be treated differently. They are not simply AI-specific complements. They are pre-existing essential sectors with constrained supply and institutional rents. The AI connection is that dividends, wage gains, or productivity benefits can be captured by landlords, insurers, healthcare systems, and credentialed access points. If the scarce goods of life remain rent-extracting, AI abundance can flow upward even when machine cognition is cheap.
The ownership regime may also be more complex than "workers versus private capital." Advanced AI may be controlled by a hybrid of frontier firms, national-security agencies, regulated utilities, pension funds, sovereign wealth funds, public-private national champions, professional bureaucracies, and cloud platforms. Corporate concentration and state-security concentration may overlap. The danger is not only private monopoly. It is concentrated control of strategic intelligence infrastructure, whether corporate, state, or public-private.
The democratic question is therefore not whether intelligence becomes abundant in the abstract. It is whether ordinary people have rights, ownership, access, and bargaining power over the systems through which intelligence becomes effective.
IX. Political Power and the Redistribution Paradox
Work is not only a way to earn money. In modern democratic capitalism, employment provides income, identity, status, healthcare, routine, mobility, and political leverage. But the source of labor power has never been economic necessity alone.
Necessary workers can still be powerless. Agricultural labor was indispensable for millennia and often had little bargaining power. Domestic workers, care workers, and service workers may be socially necessary while remaining underpaid. Labor power emerges when necessity becomes organized: when workers are concentrated, coordinated, legally protected, represented, and capable of disrupting production or politics. If AI reduces the necessity of broad human labor while also fragmenting work into contractors, platforms, residual tasks, personal micro-enterprises, and AI-supervised roles, then the organizational basis of labor power may weaken. But political power does not arise only from labor. Citizens retain leverage as voters, consumers, tenants, taxpayers, and sources of political legitimacy. States may redistribute even when recipients are not economically indispensable. The redistribution paradox should therefore be stated carefully. If labor loses economic leverage, it may become harder for workers to force redistribution through workplace power. But democratic leverage may still exist through elections, social movements, legitimacy crises, public finance, consumer politics, and state interest in stability.
The paradox remains serious because the policies needed to stabilize a post-labor economy require political power from people whose economic power may be weakening. Pre-distribution is not an escape from this paradox. It is a race against it.
The paradox also has a fiscal floor, and the civic levers above quietly rest on it. A state funded by wages and the consumption wages finance has a material stake in mass prosperity, whatever its ideology; the taxpayer is leverage even when the worker is not. If machine cognition and its infrastructure become the tax base — a handful of firms remitting what payroll across millions of workers used to remit — that stake thins. This is the resource-curse mechanism arriving home: states funded from a concentrated source rather than from their citizens' labor have historically needed those citizens less, and governed like it. Kokotajlo states the AI version plainly: people who lose their jobs lose income and political relevance together, because a government whose revenue no longer runs through its people has weaker reasons to mind them. This gives the ownership agenda of Section X a second justification that has nothing to do with income support. Citizen claims on AI capital — a dividend from an owned fund rather than an annual appropriation — do not merely replace wages. They re-thread the state's revenue interest through the population, which is what the wage-tax link did for democratic capitalism all along.
The fiscal floor has a twin. The modern democratic state was built on two dependencies it could not synthesize: revenue and soldiers. The historians of state formation, Charles Tilly first among them, tell the story of that bargain — rights traded for taxes and conscripts, the franchise expanding with the mass army and the mass tax base. Machine cognition promises to synthesize the first, and increasingly autonomous weapons the second. A state whose treasury does not run through its people's paychecks and whose security does not run through its people's bodies has lost the two material reasons that forced it, again and again, to bargain with them. What remains is legitimacy — real, but thinner than the historians' pair. This is why the ownership agenda of Section X is not welfare policy. It is a replacement bargain.
Who Does the Owning?
A race needs a runner, and I should be honest about who mine is. Three answers, in descending order of my confidence.
First, compression creates its own political agent in a way earlier displacements did not. The workers most exposed — analysts, paralegals, marketers, junior engineers, writers, designers — are the most articulate, credentialed, digitally organized class in history, concentrated in the metropolitan areas where political attention gathers. The agricultural precedent is sharper than the usual telling. Agrarian decline produced the loudest political movement of its era — the Populists nearly took the presidency in 1896, lost, and watched their planks become law a generation later — and when displaced agriculture finally converted decline into permanent subsidy, the victory went to landowners wielding an institutional lever, the malapportioned Senate, while the field labor got the road to Chicago. The sector split by ownership: this essay's thesis in miniature, hiding inside the standard counterexample. Displaced knowledge workers write the pamphlets, staff the campaigns, and run the group chats. A movement of the compressed will not lack for communications talent. What it lacks is the older leverage. Compression removes the strike — you cannot stop production you no longer perform — and knowledge workers hold no equivalent of the farm bloc's Senate. Their remaining asset is that they staff the institutions that manufacture belief: media, law, schools, administration. A class that cannot stop the economy can still stop the story. Whether that is enough to convert grievance into institutional power is exactly the open question, but the movement will not fail for lack of voice.
Second, state capacity is not the missing ingredient. In June 2026 a single Commerce Department export-control order took the most capable deployed models in the world offline within hours, three days after they shipped. Whatever else that episode proved, it settled the question of whether the state can act on this infrastructure when it wants to. Honesty about the episode cuts both ways: the capacity it proved was security-shaped — executive, instant, undeliberated — which suggests the fastest vehicle for public claims may be security nationalism rather than social democracy, at exactly the price this section ends on. What is missing is motive, and motive is what legitimacy crises supply. States redistribute not when workers are strong but when stability is threatened — and a compression decade is a stability threat with a long fuse.
Third, and honestly: the paradox may win. Section X should be read as a target to build toward while leverage exists, not a forecast of what will be built. If the window closes — if ownership concentrates faster than public claims are established — then the policies below become the demands of a weaker position, fought for on worse terrain. That is precisely the argument for treating the no-regrets tier as urgent rather than optional. The cheapest time to buy insurance is before the diagnosis.
Each of these runners is on a clock. The compressed class is most organized early, before displacement disperses it into gig work and geography. The state's material stake is largest now, while roughly four-fifths of federal revenue still arrives through individual income and payroll taxes — a share that shrinks as the tax base migrates. Even the potential ally inside capital — every firm that needs customers with money — holds only until demand decouples from households. Every source of countervailing power in this section is a wasting asset, and they all waste in the same direction. That is what the race means.
The strongest response is to build public claims before full displacement. If democratic societies establish public equity stakes, compute taxes, AI dividends, sovereign AI wealth funds, public data trusts, or public returns on subsidized infrastructure while citizens still have leverage, redistribution becomes embedded before labor weakens further.
National security complicates this. Once advanced AI appears strategically decisive, governments may treat it less like ordinary software and more like military, economic, and scientific infrastructure. The state may regulate rents, claim public returns, and build national AI capacity. It may also increase secrecy, concentrate power, protect incumbents, expand surveillance, and subordinate labor policy to strategic competition. Corporate rentier concentration and state-security concentration are not opposites. They may reinforce each other. The same firms that control compute, models, and cloud infrastructure may become state partners. The same state that could demand public equity may also shield strategic AI systems from democratic oversight. The future ownership regime may be a public-private security-industrial complex rather than a simple free market.
That is why the labor question becomes a constitutional question: who owns, governs, audits, accesses, and benefits from the infrastructure of machine intelligence?
X. Policy Under Uncertainty: Own, Open, Guarantee, Recognize
A serious policy framework should follow from the diagnosis. If rents migrate to the complements around intelligence, then democratic policy must target those complements.
The policy agenda can be organized around four imperatives: own, open, guarantee, and recognize.
The imperatives are scenario-weighted. Under augmentation, open matters most: the gains exist and the fight is over who captures them. Under compression, own and guarantee become urgent while labor still has leverage to demand them. Under substitution, all four are load-bearing. And the prescriptions carry costs the diagnosis should admit: taxing compute rents and taking public equity can slow the buildout that produces the abundance worth distributing. That tradeoff is real. The argument here is only that the cost of embedding public claims early is small next to the cost of trying to establish them late.
1. Own the bottlenecks
If labor income declines, society needs another way to distribute purchasing power. That means public claims on AI capital and its complements: compute infrastructure, energy access, model rents, strategic data, distribution platforms, and other systems through which machine intelligence becomes economically useful.
This could take the form of sovereign AI wealth funds, public equity in publicly supported AI infrastructure, compute-rent taxation, model-rent taxation, data-trust revenues, licensing conditions, public procurement stakes, or public ownership of strategic infrastructure. If the state provides subsidies, land, energy access, procurement contracts, security protections, or legal privileges, it can demand public returns.
An AI dividend funded only by annual taxation may be politically fragile. A durable fund that owns claims on productive assets is stronger. In a world where ownership matters more than wages, democratic societies need citizens to hold capital claims directly or through public institutions.
2. Open the bottlenecks
Ownership is not enough if private gates capture the gains. Anti-rent policy must prevent essential systems from becoming toll roads. That means interoperability, data portability, antitrust enforcement, limits on exclusive compute contracts, open standards, platform access rules, common-carrier-like obligations for essential AI infrastructure, and constraints on vertical integration where it locks users into a single stack. Housing and healthcare belong here too. A dividend in a housing-constrained society can become a subsidy to landlords. Public compute in a closed-platform society can become a subsidy to platform owners. AI abundance becomes social abundance only if rent-extracting sectors do not capture the surplus.
3. Guarantee agentic access
Public compute should be understood as one part of public agentic infrastructure: models, tools, data permissions, identity systems, APIs, audit logs, legal delegation frameworks, and public digital institutions that let citizens use machine intelligence to act in the world.
In a simple version, every person should have access to a capable public agent for taxes, benefits, healthcare navigation, legal letters, education, creative work, financial planning, accessibility, and life administration. This is the public-library version of machine intelligence: baseline cognitive agency available to everyone.
But access must include rights. A personal agent is powerful only if it can act. It needs permission to retrieve data, file forms, transact, communicate with institutions, negotiate, appeal decisions, schedule services, and represent a person within defined legal limits.
A democratic AI society may therefore require a right to machine representation: the right for individuals to delegate limited authority to AI agents that can interact with firms, platforms, agencies, insurers, landlords, and institutions on their behalf. If corporations and governments have agents while citizens do not, asymmetry deepens.
Representation also requires loyalty, and this is where Section IX's remaining lever — the vote — connects back to infrastructure. If daily advice, news digestion, and civic information all run through agents owned by the firms whose power is in question, then the channel meant to discipline concentrated power is operated by it, and a bias too subtle for any user to detect, multiplied across a population, is a constitutional problem rather than a product flaw. A right to machine representation therefore needs a fiduciary condition: an agent that represents a person must be bound to that person's interest, auditable against it, and barred from undisclosed optimization for its operator. The public agent is where that standard should be set first. Its loyalty, not just its availability, is the point of building it.
Civic compute funds and cooperative agent institutions are speculative designs, not the core income mechanism. Raw compute without data, permissions, distribution, trust, and management is weak. The central goal is usable agency.
4. Recognize contribution beyond wage labor
If work no longer functions as the default distributor of role, routine, status, and recognition, society must build institutions that let people contribute without being forced into artificial scarcity or meaningless jobs. This includes civic service, care institutions, ecological restoration, lifelong learning, and democratic participation. These are not decorative extras. They are role-distributing institutions.
The goal is not to preserve bad jobs for meaning's sake. It is to create socially recognized ways for people to be responsible, useful, challenged, and needed.
Contribution-Weighted Claims
Recognition without material teeth repeats the oldest failure in this domain. Care work has always been recognized in the applause sense — essential, praised, and underpaid. If the fourth imperative floats free of the first, it becomes ceremony. The repair is to couple them: recognized contribution should carry augmented claims on the systems built under Own. A sovereign AI fund can pay a baseline dividend to all and a contribution-weighted supplement to those doing the work societies say they value — care, teaching, civic service, ecological restoration, training the next generation of practitioners in fields machines have thinned. This is not a wage, and it should not be means-tested into a bureaucratic gauntlet. It is a share: the mechanism by which "we need this done" stops being a sentiment and becomes a claim. A post-labor society that pays for what it praises will get more of it. One that merely praises will get burnout and applause.
These policies should be staged. The no-regrets tier is cheap under every scenario and should be built now: interoperability, data portability, liability rules, auditability, worker consultation over deployment, and public returns on publicly subsidized infrastructure. The compression stage adds sectoral bargaining, apprenticeship redesign, reduced hours, wage insurance, and public-service job creation. The substitution stage escalates the imperatives to their strong forms: social dividends, universal services, sovereign AI funds, and machine representation recognized in law.
Staging carries an assumption that should be named: that the stages announce themselves — that visible compression arrives early enough, and legibly enough, to create the political demand for compression-stage policy while leverage remains. Under the takeoff route of Section IV, that assumption fails: the quiet period is precisely when these demands sound alarmist, and the loud period arrives after the leverage is gone. The corollary is hard but simple. The no-regrets tier is not merely the first tier. It is the only tier guaranteed a political window, and it deserves the urgency the later tiers may never get the chance to claim.
If automation is competitively rewarded, moral appeals to "keep humans in the loop" will be weak unless backed by countervailing institutions. Human-centered production requires power: unions, professional standards, liability rules, public procurement requirements, customer rights to human review, antitrust, interoperability, public ownership, and democratic oversight. Otherwise the market's default is not to preserve human roles, but to minimize dependence on them where doing so is profitable.
Conclusion: The Crisis of Economic Necessity
Advanced AI would not make humans meaningless. Human beings would still love, suffer, create, care, play, compete, worship, argue, teach, and build communities. The worth of human life is not determined by market demand.
But modern society is not organized around that truth. It is organized around work. Democratic capitalism assumes that most adults can sell useful labor in exchange for security and dignity. If machine cognition becomes abundant and cheap, that assumption weakens.
AI capability creates possibility. Capitalist competition turns possibility into pressure. Ownership determines who captures the gains. Institutions determine whether ordinary people retain leverage.
Work has been a bad but functional meaning machine. If it weakens, the goal should not be to preserve artificial jobs for their own sake, but to build better institutions for income, agency, recognition, and free contribution under conditions of dignity. The essay's central claim is not that intelligence itself becomes both cheap and toll-gated. It is that routine cognition may become cheap while the scarce complements around useful cognition become the new toll gates. Labor devaluation comes from cheap baseline cognition. Political dependency comes from enclosure of the systems through which cognition becomes useful.
The optimistic future is possible. AI could raise living standards, accelerate science, improve education, expand access to services, and free human beings from drudgery. It could repair ladders, enable micro-enterprises, lower prices, and make individuals more capable. But abundance is not automatically shared, and agency is not automatically equal.
If AI surplus flows mainly to the owners of compute, energy, platforms, housing, healthcare, and state-security systems, then a richer society could also become a less equal and less democratic one.
The democratic alternative is not merely a stipend society in which machines work, owners profit, and the state sends checks. It is a society in which people hold claims on AI capital, receive returns from machine productivity, possess usable access to productive intelligence, and retain power over the bottlenecks around that intelligence.
Income keeps people alive. Agency keeps them capable. Recognition gives them a place in the world. Anti-rent power prevents the gains from being captured.
Written from the June 2026 frontier, the claim is not that the future is settled. It is that the ownership structure of machine intelligence is being built before its social contract exists.
The future after cognitive scarcity will be decided not only by what machines can do, but by who owns them, who governs them, who receives their surplus, who can access their intelligence, who controls the scarce complements around them, and whether ordinary people retain power, dignity, and recognized roles in a world that may no longer need their work in the same way.
Selected Bibliography and Intellectual Lineage
This essay is synthetic rather than a formal literature review. The following bibliography is a scaffold for a publication-grade version, which should verify all dates, editions, URLs, and claim-level citations.
AGI, ASI, and Frontier AI Trajectories
OpenAI. "OpenAI Charter." 2018.
OpenAI. "Planning for AGI and Beyond." 2023.
OpenAI. "Introducing GPT-5.5." 2026.
OpenAI. "GPT-5.5 System Card." 2026.
Anthropic. "When AI Builds Itself." 2026.
AI 2027. "Takeoff Forecast." 2025-2026.
AI Futures Project (Kokotajlo, Daniel, et al.). "AI 2040: Plan A." July 2026.
Kokotajlo, Daniel. Interview on The Diary of a CEO. July 2026.
OpenAI, Oracle, and SoftBank. Stargate infrastructure announcements and updates. 2025-2026.
U.S. Department of Commerce. June 2026 suspension order on frontier-model access, and related reporting.
Genewein, Tim, et al. "From AGI to ASI." 2026.
Karpathy, Andrej. "Jagged Intelligence" and related "Year in Review 2025" remarks.
Labor Economics, Automation, and Wages
Autor, David H. "The 'Task Approach' to Labor Markets: An Overview." 2013.
Autor, David H. "Why Are There Still So Many Jobs? The History and Future of Workplace Automation." 2015.
Acemoglu, Daron, and Pascual Restrepo. "Artificial Intelligence, Automation, and Work." 2018.
Bessen, James. "Learning by Doing: The Real Connection between Innovation, Wages, and Wealth." 2015.
Bessen, James. "Technology and Learning by Factory Workers: The Stretch-Out at Lowell, 1842." 2003.
Acemoglu, Daron, and Pascual Restrepo. "Automation and New Tasks: How Technology Displaces and Reinstates Labor." 2019.
Acemoglu, Daron. Recent skeptical macroeconomic work on AI productivity and near-term aggregate effects.
Brynjolfsson, Erik, Daniel Rock, and Chad Syverson. "The Productivity J-Curve." 2021.
Brynjolfsson, Erik. Work on AI, complementarity, and the "Turing Trap."
Korinek, Anton. Work on AI, wages, labor displacement, and economic policy under advanced AI.
Baumol, William J., and William G. Bowen. "Performing Arts: The Economic Dilemma." 1966.
Empirical AI Productivity Evidence
Brynjolfsson, Erik, Danielle Li, and Lindsey Raymond. "Generative AI at Work." NBER Working Paper No. 31161. 2023.
Dell'Acqua, Fabrizio, Edward McFowland III, Ethan Mollick, Hila Lifshitz-Assaf, Katherine Kellogg, Saran Rajendran, Lisa Krayer, Francois Candelon, and Karim R. Lakhani. "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality." 2023.
Peng, Sida, Eirini Kalliamvakou, Peter Cihon, and Mert Demirer. "The Impact of AI on Developer Productivity: Evidence from GitHub Copilot." 2023.
METR. "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity." 2025.
Stanford Institute for Human-Centered AI. "Artificial Intelligence Index Report 2026." 2026.
OpenAI. GDPval and related economically valuable work evaluations.
Supervisory Ratios and Automation Precedents
Olsen, Dan R., and Stephen Bart Wood. "Fan-Out: Measuring Human Control of Multiple Robots." 2004.
METR. "Measuring AI Ability to Complete Long Tasks." 2025. "Time Horizon 1.1" update. 2026.
Anthropic. Engineering disclosures on multi-agent Claude sessions and model-written code share. 2026.
Waymo. Remote-operations and fleet-oversight disclosures. 2026.
U.S. Nuclear Regulatory Commission. NuScale control-room staffing approval materials. 2023.
President's Task Force on Aircraft Crew Complement. Report. 1981.
Federal Aviation Administration. Part 108 beyond-visual-line-of-sight rulemaking. 2025-2026.
U.S. Air Force. Remotely piloted aircraft staffing and enterprise personnel materials.
Political Economy, Rent, and Ownership
Varoufakis, Yanis. "Technofeudalism: What Killed Capitalism." 2023.
Durand, Cedric. "Techno-feodalisme: Critique de l'economie numerique." 2020.
Zuboff, Shoshana. "The Age of Surveillance Capitalism." 2019.
Srnicek, Nick. "Platform Capitalism." 2016.
Lanier, Jaron. Work on data dignity, data dividends, and human contributions to digital systems.
Tilly, Charles. "Coercion, Capital, and European States, AD 990-1992." 1992.
Ross, Michael L. "Does Oil Hinder Democracy?" 2001.
Alaska Permanent Fund materials and sovereign wealth fund models.
Public Compute, Agency, and AI Governance
Public compute proposals and AI public-option discussions.
Data portability and interoperability policy.
AI accountability, liability, and auditability law.
Agentic AI governance and delegation frameworks.
Labor unions, sectoral bargaining, works councils, data cooperatives, and compute cooperatives.
Meaning, Work, and Recognition
Marx, Karl. "Economic and Philosophic Manuscripts of 1844."
Arendt, Hannah. "The Human Condition." 1958.
Durkheim, Emile. "The Division of Labor in Society." 1893.
Durkheim, Emile. "Suicide." 1897.
Honneth, Axel. "The Struggle for Recognition." 1995.
Frankl, Viktor. "Man's Search for Meaning." 1946.
Contemporary post-work, degrowth, and leisure-society debates.
Historical Analogies
Catalini, Christian. "Babysitting the Slop." Forbes, March 2026. (Prior use of Bessen's loom history as an analogy for AI-agent supervision; the stretch-out framing here shares that anchor.)
Agricultural mechanization and the decline of farm employment.
Electrification and regulated utility infrastructure.
Computerization and software automation.
Industrial labor organization, unionization, and sectoral bargaining.