The Civilization That Cannot Explain Itself

Every institution can still account for itself. None of them answers for what they are doing together.

The dashboards are still green.

Output is growing, the models keep improving, and capital keeps pouring into the buildings that run them. The first essay in this series closed on that image: the lights on, the shelves full, the dashboards turning green while the bargain that made work legitimate came under strain. Eleven essays later the dashboards have not changed color. What they were never built to show is the subject of this last one.

Go back to the junior analyst that first essay imagined. She did what the institutions around her recommended. She took the degree, carried the debt, and prepared for the kind of first-year work that used to pay for an apprenticeship in judgment: the summaries, the first drafts, the market scans a senior colleague would mark up. The firm she applied to can explain why it now hires fewer people like her, since its senior staff produce more with better tools. The university can explain its tuition. The software that screened her application can explain its ranking rules, and the agency that would hear her complaint can explain its appeal procedure. Each account is coherent on its own terms. Nobody is responsible for their sum, which is the condition she actually lives in.

That is what this series has meant by a civilization that cannot explain itself. Its institutions can still justify their own conduct. They cannot account for the condition they produce together, and none of them is obliged to try. Elected government is supposed to answer for the whole, and in principle it does. In practice the authority to decide, the information needed to see the pattern and the remedies a person can use are spread across the firm, the school, the vendor and the agency, so the institution with the general mandate rarely holds the facts or the levers when her case arrives.

Explaining, in this sense, has little to do with storytelling. A better story would not repair anything. An explanation that holds a society together is one a person can test: what she is entitled to, who decided the matter that affected her, whether she can challenge it, and whether anyone can change the system that produced it. For about a century the wage answered those questions well enough for enough people. The argument of these twelve essays is that the institutions which carried the answers are being hollowed out at particular points, and that a replacement is being drafted now by parties with a strong interest in its terms.

Six Machines, Six Promises

The wage was the first (E3). It gave an ordinary person a recognized claim: I contributed, so I am owed a place. It also made unequal ownership tolerable, because someone who did not own the factory could still expect employment, some bargaining power and a share in what the factory made. It was never universal. People doing unpaid care contributed without receiving its protections, and that exclusion matters for whatever might replace it. AI does not need to abolish jobs to weaken the claim. It needs only to make some kinds of contribution less scarce at the point where people enter.

The evidence on that point is real and mixed. Stanford’s payroll research, revised in August, puts employment of 22-to-25-year-olds in the most AI-exposed occupations about 19 percent below where it would be had it kept pace with less-exposed peers of the same age, with the adjustment running through fewer hires rather than more layoffs; the authors present the patterns as descriptive rather than causal and report no economy-wide displacement. [1] Yale’s Budget Lab, updated in October, still finds no clear AI footprint in its labor-market measures, [2] and a study of recent college graduates finds no unusual rise in their unemployment this summer. [3] If hiring into exposed occupations recovers as firms redesign junior work around the new tools, the claim that the wage is losing force at the point of entry would weaken, and the series would have mistaken a pause for a withdrawal.

The aggregate numbers cannot settle that question in either direction, because they measure something else. Labor’s share of nonfarm business output fell to 52.8 percent in the second quarter, the lowest since the series began in 1947, though how much of the decline survives a different treatment of taxes is disputed. [4] A falling share would not prove the entry argument, and a steady one would not refute it. The claim concerns who gets through the door: total pay can hold up while it concentrates among people already inside, and the route in can narrow without moving the aggregate at all.

History counsels patience here. About six in ten American jobs in 2018 were in specialties that did not exist in 1940, [5] and new work has repeatedly renewed the demand for labor after old work disappeared. The open question is whether the new work will supply income, formation and bargaining power to the people losing the older routes, or only to the people already established when it arrives.

The market was the second (E2). Its justification was the chance to compete: a seller might lose, but she could make an offer. As eligibility moves upstream into payment rails, procurement filters, ranking rules and agents buying on someone’s behalf, the question becomes whether she reaches the contest at all. A better price cannot help a firm the buyer’s agent never considered. The claim would weaken if agentic buying keeps the contest open, with sellers able to see why they were passed over and to dispute it.

Capital was the third (E6). It justified its returns by building and bearing risk, and those remain real contributions. The difficulty begins when the reward for building a useful system becomes a durable power to set the terms on which others work. That is why compute behaves more like an estate than a factory (E5), and why the firms and governments that depend on it look more like tenants than customers (E4): a subscription can carry obligations that resemble governance while giving the subscriber little say over the rules. Cheaper and open models would answer this to the extent that they give users somewhere practical to go, beyond weights they can download.

The state was the fourth (E9). It justified its authority by the answerability of its decisions. As judgment moves into procured systems it does not fully control, the state keeps the liability without keeping the authorship; it becomes the allocation state, answering for outcomes it did not design. The same essay named what would let it answer again: the capacity to interpret what a system does, to override a particular outcome, to replace a provider before a crisis forces the question, and to shape a system before buying it. Where agencies keep those capacities, the diagnosis does not hold for them.

The university was the fifth (E10). It sold a route into a professional beginning. The credential still clears the gate, but the junior work through which beginnings turned into judgment is being compressed, and whether supervising machine output can form judgment the way doing the work once did remains unknown. One early experiment points somewhere uncomfortable. In a three-month randomized trial with 133 patent lawyers, AI access improved drafting quality for juniors and seniors alike while they used it, but when the tool was taken away and they reviewed and redlined an application unaided, only the seniors kept a clear gain; the juniors’ results spread out, with more strong results and more poor ones. [6] If new junior roles appear that allow mistakes at low stakes and correction at close range, the ladder will have been rebuilt.

The standard was the sixth (E11). Technical standards offered reasons to trust systems too complicated for any single buyer to inspect. The trouble is rarely a missing authorization at the origin; the leading American AI framework was written at Congress’s request. It lies in the authority a framework accumulates as procurement officers, auditors, insurers and legislatures build on it, until changing it would require all of them to move at once. Colorado showed both halves of that in May, when it repealed its framework defense and replaced it with narrower duties of notice, explanation, human review and documentation, the kind of duties the framework had been built to help discharge. Whether firms keep running framework-based procedures to meet them once the new law takes effect in January is the test.

The six machines share a failure. Each still runs and produces its outputs; what is shrinking is the ability to redeem the promise each one made. A credential still signals selection. A vendor can still describe its terms as voluntary. An agency still accepts appeals. Whether those forms deliver a beginning, a usable alternative or a remedy is a separate matter, and it is the matter these essays have been pressing.

What Made the Old Account Hold

It would be easy to hear all this as a communications problem, as if the new order needed only better salesmanship. The wage’s account never worked that way. A union contract made part of the worker’s claim enforceable against an employer. Social insurance carried protection past the next paycheck. Public education supplied a route into work that a household did not have to buy alone. Courts and elections gave citizens ways, however uneven, to contest the institutions that governed them. The promise of membership was credible where people could press it against someone who would rather not have honored it.

No single institution designed that settlement. Workers organized, legislatures passed laws, employers conceded under pressure, and the wage came to carry a moral account because institutions had been forced to make the account true. That history also limits nostalgia. The old settlement often failed people outside its favored kinds of employment, and restoring its exclusions would not restore its legitimacy.

The same test applies to money that arrives without a job attached. A payment can be an entitlement secured by law, a share in a collectively held asset, emergency relief, or a favor that can be withdrawn. The amount matters, and so does the relationship it creates between the recipient and whoever controls it. A transfer by itself confers little standing. A defined entitlement, terms that change only through a public procedure, and a remedy when it is denied can confer a good deal. The difference lies in the institutions around the payment.

What Populism Reads

Populism is often treated as a failure of understanding: people too angry or too poorly informed to grasp a complex system. That reading protects the system by locating the problem in the people. Another reading, set out in an earlier essay on why the machine still working is the problem, treats populism partly as a measurement of the distance between the official account and lived experience. It is a crude instrument. Political entrepreneurs exaggerate grievances, point them at outsiders, and mobilize people whose own position is secure; cultural conflict and the supply of parties willing to exploit it matter as much as economics. Populist anger regularly attaches a real grievance to the wrong cause. The distance it registers can still be real.

AI adds a particular experience to societies already divided over housing, migration, credentials and public services: being told the economy is advancing while the route through which you expected to join it becomes harder to find. That experience can change politics before the economic facts are settled. In experiments published in PNAS in January, British and American participants who read that AI would replace human labor reported lower trust in democratic institutions and less willingness to engage politically over AI’s future. [7] The study measured a framing, and says nothing about what losing a job does. The response it found was withdrawal.

That finding bears on the strongest objection to this whole line of argument. If rulers and firms no longer need much labor, perhaps they no longer need much consent, and the gap between output and explanation will simply be tolerated by people with less and less leverage. The record so far supports a narrower worry. Consent still carries prices in the democracies at the frontier: elections, courts, permitting fights, procurement decisions, the votes of people asked to accept the physical plant the buildout requires. What can erode is the willingness to use them. A decline in populist support would therefore prove little about legitimacy, since it might mean people had stopped expecting politics to help. Recovering trust together with rising participation in the decisions that govern AI would tell against this diagnosis. A quieter electorate would not.

The Dangerous Interval

The dangerous interval is the stretch in which old claims lose force before credible replacements acquire it. The industrial transition that produced the wage settlement took generations, and its interval was not gentle: land, guild and parish lost their hold long before union, social insurance and the citizen-worker were assembled into something people could rely on. Nothing guarantees a comparable length or a comparable violence this time. The interval can contain reform as easily as resentment.

What makes it dangerous is the position of the people inside it. They have to make irreversible decisions about debt, training, children and where to live while institutions are still debating whether a transition is happening at all. A firm that trims its intake today books the saving now, while the cost of a missing cohort of experience arrives later and somewhere else. A growing economy could absorb that loss. It could also hide it for years. Transfers, platforms and the other containment arrangements traced earlier (E7) can quiet the political expression of the gap without closing it, and a gap that is managed rather than closed leaves time for the next account to be written by whoever is ready first.

Who Is Writing the Next Bargain

That next account is no longer hypothetical. Drafts are on the table, from a senator and from the companies building the systems.

In June, Senator Bernie Sanders introduced the American AI Sovereign Wealth Fund Act, which would impose a one-time levy, paid in stock, on the largest AI companies, place the shares in a public fund run by a Senate-confirmed commission, and direct the fund’s earnings toward payments to Americans. His justification is the one this series has called the inheritance grammar, first set out when the wage broke (E3) and carried through capital’s thinning justification (E6) and the tax state’s search for new ground (E8). AI, the senator argues, rests on “the collective knowledge of humanity and the creative work of tens of millions of people,” and “when a public resource generates wealth, the public should share in that wealth.” [8] In April, OpenAI published an industrial-policy paper proposing a public wealth fund whose returns “could be distributed directly to citizens,” leaving it to “policymakers and AI companies” to decide how to seed it. [9] In August it funded fourteen outside projects on ownership, benefits and taxation, stating that “these choices should not be made by technology companies alone.” [10] On governance, Anthropic’s chief executive has proposed embedded outside evaluators and industry coordination while legislation develops, the proposal the previous essay examined. None of these is law, and none yet establishes an entitlement.

A proposal is not discredited by its sponsor’s commercial interest, and some of these would be real concessions. But whoever supplies the first serious draft gets to define the problem, choose which options receive attention and set what a reasonable settlement looks like. The question the previous essay asked of technical standards returns here at a larger scale. The institution supplying the capability can also supply the vocabulary in which everyone else discusses what they are owed for accommodating it. Officials need something usable, adoption breeds familiarity, and familiarity narrows what later debate treats as practical. None of that is inevitable. Governments can reject drafts, rewrite them, compel disclosure and build their own expertise. What remains scarce is a way to turn a proposal into a bargain its authors cannot later revise on their own.

The inheritance argument has real force and real limits. AI depends on accumulated science, language, law, infrastructure and culture, and also on firms that financed and assembled systems that would not otherwise exist; their contributions warrant returns without exhausting the sources of value. Shared origins cannot calculate a fair dividend. A national fund cannot simply claim knowledge produced around the world, and the boundaries of membership, the size of the public claim and the institutions administering it are political decisions that no inventory of training data can settle. What the argument does supply is a reason for people to claim a share beyond the shrinking market value of their own labor, and that reason grows more urgent as the institutions collecting the returns gain control over other people’s opportunities.

A share in the proceeds is still only half a settlement. A citizen could hold a dividend and have no say in how she is classified, which systems sort her, or what counts as a harmful decision. The Sanders bill shows how much turns on that difference. Its governance is carefully built: commissioners confirmed by the Senate, a seat reserved for labor, bars on holding AI equity, and a quarterly public record of every shareholder vote. Its payment is weaker than its announcement. The bill authorizes an annual amount, five percent of the fund’s value, to be appropriated, and leaves its use, direct payments among other purposes, to whatever Congress later provides. [11] No individual acquires a claim she could enforce; the dividend would arrive, if it arrived, as a decision each Congress makes again. OpenAI’s proposal is less specified still, naming no source of capital, no governing body and no rule for distribution. [12] If a dividend is enacted with a defined entitlement and terms that change only through public procedure, that would show the danger met by deliberate reconstruction, which is a different outcome from the diagnosis proving wrong. What would count against the diagnosis is the opposite case, in which accessible careers and accountable institutions persist without any reconstruction at all.

What a Credible Account Would Need

A society cannot promise the analyst a particular profession, protect every task from change, or make her degree immune to disappointment. It can decide who pays to form the next generation of expertise, what she is owed while she looks for another route, and who must answer when a system sorts her. The old bargain let most of those costs travel inside employment. A successor would have to name them openly.

Read across the twelve essays, any draft that hopes to be believed would need at least three things. The first is a material claim that does not depend on being economically indispensable, and that recognizes the work wages never reliably paid for, the care that keeps households and communities running. The second is a usable way to contest consequential decisions, which means someone with the authority to investigate a decision and change it, and an agency that can interpret what a system did and override its result, two of the capacities the allocation-state essay named (E9). The third is public institutions able to replace the systems through which they govern and to shape them before buying them, funded by a tax base that can see where the new surplus accumulates (E8), since a settlement promising support beyond employment cannot rest indefinitely on taxing the employment it expects to weaken. These are tests for judging a draft. Each is also a decision about who carries a cost, and that decision belongs to institutions the people carrying it can still move.

The dashboards can stay green either way. They record what the system produces. They cannot record whether the analyst who received the rejection has anyone to ask why, or anything she is entitled to ask for.

Notes

[1] Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, “No Widespread Displacement, but the AI Employment Gap for Young Workers Has Widened to 19%,” Stanford Digital Economy Lab, August 12, 2026. https://digitaleconomy.stanford.edu/news/canariesaug26/. ADP payroll data through June 2026; the 19 percent is a relative shortfall against same-age workers in less-exposed occupations, not an estimate of jobs eliminated by AI. The authors describe the results as descriptive patterns, not causal estimates, and find the adjustment operating mainly through reduced hiring. This is the revised version of the payroll study whose earlier figure opened this series.

[2] The Budget Lab at Yale, “Tracking the Impact of AI on the Labor Market,” updated October 14, 2026. https://budgetlab.yale.edu/research/tracking-impact-ai-labor-market. Its synthetic difference-in-differences analysis “does not yet clearly indicate an AI-related labor market footprint.” Stanford and Yale use different data and methods, and the two findings are not a contest in which one must be wrong.

[3] Robert W. Fairlie and Jane Wu, “The Early Impacts of AI on Employment among Recent College Graduates,” NBER Working Paper 35796, September 2026. https://www.nber.org/papers/w35796. Using Current Population Survey microdata, the authors find that recent graduates’ unemployment did not spike in summer 2026 relative to earlier summers or to comparison groups.

[4] U.S. Bureau of Labor Statistics, “Labor share at its lowest level, 52.8 percent, in second quarter 2026,” The Economics Daily, September 15, 2026. https://www.bls.gov/opub/ted/2026/labor-share-at-its-lowest-level-52-8-percent-in-second-quarter-2026.htm. Nonfarm business sector; the series begins in 1947. BLS does not attribute the decline to AI. On measurement: Joel Prakken, “AI and Labor’s Share of Income,” Haver Analytics, July 23, 2026, https://www.haver.com/articles/ai-and-labor-s-share-of-income, argues that “there has been no decline in labor’s share of income when correctly measured” once taxes on production and imports are excluded from the denominator, an adjustment BLS does not make. His analysis was published before the second-quarter figure, so it qualifies the measure rather than rebutting that reading.

[5] David Autor, Caroline Chin, Anna Salomons and Bryan Seegmiller, “New Frontiers: The Origins and Content of New Work, 1940-2018,” Quarterly Journal of Economics 139, no. 3 (2024): 1399-1465. https://doi.org/10.1093/qje/qjae008. The estimate concerns new job specialties within occupational categories, not a claim that most broad occupations are wholly new.

[6] David Autor, Tanya Rodchenko, Josh Martin, Zanna Iscenko, Scott Strand, David Pearl and Melissa Ferere, “Does AI Assistance Enhance or Erode Expertise? Evidence from a Three-Month Field Experiment in Patent Drafting,” NBER Working Paper 35720, September 2026. https://www.nber.org/papers/w35720. Eleven U.S. intellectual-property firms. Junior lawyers showed no average gain on the later unaided review-and-redlining task, with outcomes bifurcating; the authors suggest foundational expertise may be a prerequisite for building durable skill from AI-assisted practice. One trial in one profession.

[7] Armin Granulo, Andreas Raff and Christoph Fuchs, “Perceiving AI as Labor-Replacing Reduces Democratic Legitimacy and Political Engagement,” Proceedings of the National Academy of Sciences 123, no. 4 (2026), e2523508123. https://doi.org/10.1073/pnas.2523508123. The experiments, one in the United Kingdom and one in the United States, compare responses to alternative framings of AI; they do not measure the consequences of actual displacement. The paper also reports survey evidence from 38 European countries.

[8] Office of Senator Bernie Sanders, “Sanders Introduces Legislation to Create $7 Trillion AI Sovereign Wealth Fund,” June 18, 2026. https://www.sanders.senate.gov/press-releases/news-sanders-introduces-legislation-to-create-7-trillion-ai-sovereign-wealth-fund/.

[9] OpenAI, “Industrial Policy for the Intelligence Age: Ideas to Keep People First,” April 2026. https://cdn.openai.com/pdf/561e7512-253e-424b-9734-ef4098440601/Industrial%20Policy%20for%20the%20Intelligence%20Age.pdf. The Public Wealth Fund passage states that “Policymakers and AI companies should work together to determine how to best seed the Fund” and that “Returns from the Fund could be distributed directly to citizens.” A proposal, not an enacted program.

[10] OpenAI, “New Policy Ideas for the Intelligence Age,” August 17, 2026. https://openai.com/index/new-policy-ideas-for-the-intelligence-age. Grants to fourteen independent projects, $1 million in total funding plus up to $1 million in API credits. The page states that “in democratic societies, they should be shaped through democratic institutions and public debate.”

[11] American A.I. Sovereign Wealth Fund Act, S. 4825, 119th Congress (2026), bill text as posted by Senator Sanders. https://www.sanders.senate.gov/wp-content/uploads/AmericanAIWealthFundTextv618.pdf. Proposed 26 U.S.C. 9512(d) authorizes to be appropriated, for each fiscal year, an amount equal to 5 percent of the fund’s average market value, net of administrative costs, for direct payments and other purposes “as provided in appropriations Act or law.” Commission: Sec. 9512(b)(2), seven commissioners appointed with Senate advice and consent, including a representative of labor interests, with equity, blind-trust, recusal and post-service restrictions; Sec. 9512(b)(2)(B)(v) requires at least quarterly public disclosure of how the commission voted the fund’s shares.

[12] OpenAI, Industrial Policy for the Intelligence Age, note [9] (the Public Wealth Fund passage, which leaves seeding, governance and distribution to later decision).