> ## Content Index > Fetch the complete content index at: https://terawatttimes.org/llms.txt > Use this file to discover other available public pages before exploring further. # Shared-Output Governance: Productive Gains and the Institutions of Realistic Choice - URL: https://terawatttimes.org/ssc-5-shared-output-governance/ - Published: 2026-10-13T14:00:32.000Z - Updated: 2026-10-13T14:00:32.000Z - Description: Shared-Output Governance explains how AI-driven productive gains become wider social choice through three institutional levers: stable claims on output, capability reconstruction, and peaceful correction. It also shows why adjustment speed, apprenticeship, and review matter. - Author: Alex Yang Liu - Tags: Finance | Policy ## Abstract As artificial intelligence and automation raise productive capacity, the institutions that connect production to people decide whether the gains widen choice or narrow it. This paper derives those institutions from the mechanisms through which systemic survival compression forms and deepens, and it proposes that institutions reduce compression through three reachable causal levers: stable claims on output, capability reconstruction and peaceful correction. The argument begins with intelligent substitution, whose social effect is governed by the relation among four adjustment speeds, namely the rate of substitution, the formation of new tasks, the reconstruction of personal capability, and the adjustment of claims and institutions. Evidence from the United States and Denmark shows exposure and effect moving independently. Stable claims on output then combine labour, capital, public-service, cash and time channels. Capability reconstruction requires income during transition, recognised credentials, practice, entry routes and portable protection, and it must also preserve the junior work through which expertise forms, an apprenticeship layer that AI can remove while output holds. Peaceful correction covers auditing, explanation, appeal and periodic review, including review of the vendor lock-in that places the production-control problem inside the welfare state. Each lever carries an observable failure mode. Two public inquiries, into Australia's Robodebt scheme and the Dutch childcare benefits affair, show what slow correction costs, and the 2016 Alaska dividend cut shows the cost of lacking a design principle that separates stable entitlement from adjustable parameters. The paper closes by showing how capability reconstruction acts back on capability reproduction, the process through which productive abilities are maintained or lost. **Keywords:** shared-output governance; artificial intelligence and employment; automation; intelligent substitution; social protection; basic income; sovereign wealth fund; workforce reskilling; apprenticeship; algorithmic accountability; labour income share; systemic survival compression ## 1 Introduction Artificial intelligence and automation are raising what economies can produce. Whether they also widen what people and states can choose depends on the relationships through which production is organised and controlled. This paper uses the concept of **systemic survival compression** for the adverse case: the degree to which an asymmetric production or control relationship narrows the practical space in which a dependent party can maintain basic functioning, reproduce its own capabilities, obtain realistic alternatives, claim a share of productive output and correct adverse rules through peaceful means. At the domestic scale the dependent party is a worker or a household, and the relationship runs through employment, platforms and the institutions that distribute output. Responses to compression often share a pattern. Debt lets spending run ahead of income. Asset appreciation lifts measured wealth for those who hold assets. Export expansion carries a shortfall of domestic demand abroad, capacity closure trims visible excess supply, suppression lowers the visible expression of grievance, conflict reorganises coalitions while destroying productive assets, and decoupling reduces cross-border contact. Each of these reduces visible pressure in a particular place or period, and each postpones, conceals, replaces or transfers the burden while the mechanism that created it stays in place. Durable reduction in compression requires improvement in four variables. The first is **capability reproduction**, the process through which productive, occupational, technical and institutional abilities are maintained over time; the others are realistic substitution, claims on output and peaceful correction. That principle defines the task of governance. Concentration and asymmetric dependence arise from technology, scale and trade, and institutions have little purchase on them. What institutions can reach are the variables that decide how far a structural position converts into narrowed choice. Sections 2 to 4 set out the domestic scale of compression under intelligent substitution, and Section 5 defines shared-output governance. Sections 6 to 9 develop its three pillars, each with an observable failure mode, together with the design principle that joins them. Section 10 places the framework among existing governance accounts, and Section 11 concludes. ## 2 Intelligent Substitution **Intelligent substitution** is the process through which AI systems, software platforms, robots and automated equipment replace, compress or recombine tasks previously performed by people, whether cognitive, communicative, coordinative, decision-making or physical. The task is usually the right unit of analysis, because an occupation can survive while its internal structure changes. Routine drafting, scheduling, analysis or communication can be automated while judgement and accountability stay human. A profession can keep its senior positions while losing the junior tasks that once trained future experts, and employment can hold steady while discretion, bargaining power or occupational progression weakens. Exposure and effect are different measurements. On exposure, the International Labour Organization's refined global index estimates that one in four workers worldwide is employed in an occupation with some exposure to generative AI, with clerical work the most exposed, and it stresses that exposure is more likely to transform tasks than to eliminate every exposed job at once \[1\]. OECD evidence points the same way, finding early effects in work organisation, task content, monitoring, pace, autonomy and skill requirements before they appear in aggregate employment \[2\]\[3\]. On effect, the evidence divides, and the division is informative. United States payroll data show a notable relative decline in employment for early-career workers in the most AI-exposed occupations following the wide rollout of generative AI \[4\]. Danish administrative data reach the opposite result. Linking large-scale adoption surveys to employment records across eleven exposed occupations, Humlum and Vestergaard estimate precise null effects on earnings and recorded hours at both worker and workplace level and rule out effects larger than 2 percent two years on; the nulls hold for intensive users, early adopters, workplaces with substantial investment and early-career jobs alike, and adoption is linked to occupational switching and task restructuring with no net change in hours or earnings \[5\]. A European Central Bank analysis of the United States finds aggregate employment effects muted so far, with negative effects concentrated among junior workers in highly exposed occupations, and it finds that AI widened the gap in employment growth between high-risk and low-risk occupations between 2019 and 2025 without producing wage disparities \[6\]. Two national datasets covering overlapping periods and similar occupations thus disagree about whether early-career employment has moved. That disagreement is the strongest available demonstration that exposure and effect are separate quantities, and it explains why the composition of new work carries more weight than the rate of displacement. ## 3 Four Adjustment Speeds The social consequence of intelligent substitution is set by the relation among four rates of change. **Substitution speed** is the rate at which AI replaces or reorganises existing tasks, occupations and career pathways. **New-task formation speed** is the rate at which new products, services, occupations, enterprises and human-complementary activities generate new income. **Personal capability reconstruction speed** is the time workers need to learn, move into new positions, recover income and rebuild an occupational path; it is measured by the time from training to employment and by wage recovery after transition, since course completion records only an input. **Output-claim and institutional adjustment speed** is the rate at which wages, public services, social insurance, ownership, transfers and working-time rules respond, and the rate at which distributive rules themselves are revised. Domestic compression rises when substitution speed stays faster than the other three for a sustained period. The comparison is made in speeds because the same level of substitution produces entirely different social outcomes at different adjustment speeds, and level indicators such as substitution rates or unemployment rates cannot tell those outcomes apart. Think of a bathtub with the tap and the drain both open. The water level depends on the two flows together, and a reading of either flow alone says little about whether the tub will overflow. The four speeds depend on one another, so they have to be assessed together. New work raises the return to retraining, and income security lets workers finish retraining without losing basic functioning. Labour regulation and bargaining institutions shape whether firms use AI to replace workers, to monitor them or to raise their productive capacity, and public investment accelerates the formation of complementary industries. The present state of the ordering can be stated precisely, and it is partial. United States payroll evidence shows substitution running ahead of personal capability reconstruction in particular highly exposed occupations \[4\], Danish administrative evidence shows no such pattern \[5\], and aggregate assessments find limited effects to date \[6\]. The ordering therefore holds for particular occupational settings and remains unestablished across high-income economies as a class. Institutional adjustment speed has no comparable cross-national measurement and does not yet enter the ordering, while new-task formation is the least determined of the four and the variable on which the distribution of outcomes over the next decade principally turns. The four speeds are therefore measured through separate indicators and selected cases, and reported side by side as a profile. ![](https://terawatttimes.org/content/images/2026/09/Fig-1-4.webp) Figure 1\. Four Adjustment Speeds Under Intelligent Substitution ## 4 High Substitution With Low Compression, and the Reverse A society can undergo extensive intelligent substitution without severe compression. Productivity gains may raise wages, finance public services, support social dividends, shorten working time and create new enterprises, and workers leave obsolete tasks while keeping income, health care, education and social participation. The technical rate of substitution is high, and realistic choice stays broad. The reverse configuration is just as available. A limited employment shock creates acute compression when income, health coverage, housing, occupational identity and access to training all depend on a single job. International Monetary Fund staff analysis shows why the distributional result cannot be read off technical exposure \[7\]. AI may substitute for some workers and complement others. Strong complementarity with higher-income workers can widen labour-income inequality, returns to AI-related capital increase wealth inequality, and large productivity gains could still raise incomes broadly where institutions transmit them through the economy. Roughly 40 percent of global employment is exposed to generative AI, and around 60 percent in advanced economies. Earlier automation waves hit middle-skilled work hardest, whereas this time displacement risk extends to higher earners. Read against the four speeds, these findings locate where institutions act. Complementarity with some workers is a statement about new-task formation, since AI raises the value of certain human tasks and creates demand for them, and substitution for others is the substitution speed itself. Widening inequality from complementarity and from returns to AI-related capital is a statement about output claims, because productivity gains accrue to those who already hold complementary skills or capital unless claims adjust. The condition under which large gains raise incomes broadly, institutions that transmit those gains, is the fourth speed. The IMF analysis and the American and Danish data point to the same open variable, the pace at which institutions respond. The issue that decides the outcome is the relationship between technical change and the institutions that reconstruct capability and distribute output, and those institutions are the subject of the rest of this paper. ## 5 Shared-Output Governance Defined **Shared-output governance** is a system of institutions that connects productive gains to stable social claims on output, preserves realistic substitution and recovery capability, and maintains effective mechanisms of peaceful correction. It answers the practical questions an AI-intensive economy raises: who can purchase the output of intelligent production, through which channels productivity gains reach society, how individuals and states keep realistic alternatives, how concentrated control is constrained, and how institutional errors are corrected without crisis or violence. Its three pillars map onto the causal levels of compression that institutions can reach. Stable claims on output and peaceful correction are the two moderating variables, which decide how far a structural position runs into narrowed choice. Capability reconstruction reaches both transmission variables. It addresses the rising cost of replacement directly, and it addresses capability reproduction through training systems, occupational entry routes, supplier retention, maintenance competence, certification capacity and the scheduling of work itself, whose importance Section 7.1 documents. Capability reproduction converts a structural position into a trajectory. A governance design confined to claims and correction leaves that mechanism in place, so the design direction follows from the mechanism itself. The three pillars also depend on one another. Stable claims give workers the income security to retrain, which is what makes capability reconstruction usable. Capability reconstruction widens the alternatives a worker can reach, which gives correction mechanisms something to enforce, since an appeal is worth more to a person who has somewhere else to go. Correction, in turn, protects the stability of claims, because entitlements that can be revised only through reviewable procedures are harder to erode by discretion. A system that builds one pillar and neglects the others tends to see the neglected pillar undo the gains of the strong one. The framework has a scope condition. Its levers operate through formal institutions, so its reach grows with institutional coverage. Where effective social protection coverage is low, as in the low-income group documented in Section 6, the first task is building the claim channels themselves. ![](https://terawatttimes.org/content/images/2026/09/Fig-2-5.webp) Figure 2\. Shared-Output Governance: Three Causal Levers and Their Interdependence ## 6 Pillar One: Stable Claims on Output Claims on output form through five channels, and the mix counts for more than any single instrument. The labour channel works through wages, collective bargaining, profit sharing, employee ownership and productivity-linked compensation. Its economic function is to keep rising productivity connected to broad purchasing power, a function it can serve while the composition of jobs changes. The capital channel works through savings, pensions, employee funds, cooperatives, public wealth funds and public ownership of productive infrastructure, and its design has to specify ownership, management, risk, distribution and safeguards against political or private capture. Norway's Government Pension Fund Global illustrates the separation of functions such a design requires, with parliament and the finance ministry defining the framework and the central bank managing operations under a formal mandate \[8\]. The Santiago Principles set voluntary standards for governance, transparency, accountability and prudent investment across sovereign wealth funds generally \[9\]. A fund built for the AI era could hold diversified financial assets, public compute infrastructure, energy assets, intellectual property, land, data rights, or equity acquired through public support for technological development. The public-service channel operates in kind. Health care, education, housing, care, transport, digital access and basic energy services reduce the market income needed for effective social participation, and they support capability reproduction by protecting health, knowledge, mobility and time. They also raise the rate at which claims convert into consumption, because claims absorbed by essential costs never reach discretionary spending. Cash provides flexible purchasing power through social dividends, basic income, negative income taxes and targeted transfers. Its effect depends on physical supply, since extra money secures affordable housing, health care or electricity only where the relevant capacity can expand. Time, the fifth channel, converts productivity into control over life through shorter hours, paid leave and protected periods for care or education, and it gains importance as machines raise output without requiring proportionally more human labour. A claim on output is a necessary condition for consumption and an insufficient one. Claims convert into expenditure under three conditions. Income has to be stable and predictable, since uncertain income is saved. The share of essential costs borne out of pocket has to be manageable, since housing, health care, education and old-age provision absorb claims before discretionary spending begins. And a claim has to be held as a right, because a discretionary grant cannot anchor a long-term commitment such as a mortgage or a child's schooling. A design that raises claims while leaving these conditions unmet raises savings in place of demand. The public-service channel acts on the second condition, and the design principle of Section 9 acts on the first and third. A resilient system combines channels because each fails in a characteristic way. Cash leaves capability unbuilt, training leaves a person without income during transition, ownership reaches only those who already hold assets, and public services meet common needs while leaving individual preferences unmet. High returns become a distributable public rent only under four conditions: the contribution must be identifiable, an institution must be able to obtain a share, the source must be durable, and a public claim must be justified. AI returns are harder to capture than natural-resource rents, because intangible assets move across borders, returns are dispersed and excess profits may be temporary. Where the conditions fail, the other channels must carry more weight. The gap in claims is documented, and its distribution carries the analytical weight. The International Labour Organization reports that 52.4 percent of the world's population is covered by at least one cash social protection benefit, up from 42.8 percent in 2015 and past the halfway mark for the first time, leaving 47.6 percent, some 3.8 billion people, without any coverage \[10\]. Effective coverage varies sharply by income group \[10\]. Table 1\. Effective Social Protection Coverage by Income Group | Income group | Effective coverage | | ----------------------------- | -------------------------------------- | | High-income countries | 85.9 percent | | Upper-middle-income countries | 71.2 percent | | Lower-middle-income countries | 32.4 percent | | Low-income countries | 9.7 percent, barely changed since 2015 | A spread of that size is consistent with institutions shaping claims on output far more than technological exposure alone. IMF staff reach a compatible conclusion from the fiscal side, arguing that labour-saving automation could amplify job losses across both low-skill and cognitive occupations and further reduce the labour share relative to capital \[11\]. The characteristic failure of this pillar is an entitlement that remains dependent on discretionary appropriation or veto. A claim that can be withdrawn through the ordinary budget process loses the predictability that turns it into expenditure, and Section 9 documents a case. ## 7 Pillar Two: Capability Reconstruction For individuals, practical choice requires more than formal access to education. A functional system of capability reconstruction provides income during retraining, recognised credentials, opportunities to practise new skills, routes into new occupations, portable social protection and access to productive digital tools. Each closes a particular gap between a formal right to retrain and a realistic route to new work. Income lets a worker complete a course without losing basic functioning. Credentials make new skills legible to employers who cannot test them directly. Practice turns instruction into competence that holds up on the job, and entry routes turn that competence into an offer of employment. Portable protection keeps health cover and pension rights intact across the move, and digital tools give the worker the same complements as incumbents. Missing any one of these leaves the others underused. People also need some independence from a single employer or platform, which portable benefits, multiple income channels, open tools and basic public services provide. The pillar reaches the capability-reproduction variable through the touchpoints named in Section 5, and its performance shows in the time from training to employment and in wage recovery after transition. ### 7.1 The Apprenticeship Layer One mechanism deserves particular attention, because AI acts on it directly. Capability reproduction can fail when the throughput that sustains practice is interrupted, even where entrants are available. French nuclear construction documents the pattern. A state-commissioned audit of the Flamanville 3 reactor found a generalised loss of industrial expertise in a supply chain that had been dormant for more than two decades, and recommended improved training for workers, welders in particular \[12\]. Knowledge occupations are exposed to the same mechanism through their junior tasks. Drafting, first-pass analysis, document review, routine coding and scheduling are technically amenable to automation, occupational exposure is widespread \[1\], and realised employment effects remain disputed \[4\]\[5\]\[6\]. Junior tasks are output and training at the same time. A first-year analyst's rough model is a product the firm uses, and it is also the exercise through which the analyst learns to build a good one. When such tasks are automated while senior work stays human, an occupation's output can hold while the flow of work through which competence forms stops reaching people. The **apprenticeship layer**, the band of supervised, lower-stakes work through which novices become experts, can thus disappear without any fall in output. If junior-task automation outpaces the creation of alternative routes to competence, a profession will pass through a period of stable or rising output with a contracting pipeline, and the failure will become visible only when the profession tries to replace its senior cohort. The entry-level share of postings within an occupation, the ratio of supervised to unsupervised work, the time from entry to independent practice and the age structure of qualified practitioners are the leading indicators. Training policy therefore has to examine how competence forms, going beyond the skills listed in current vacancies, and the indicators show where that examination is most urgent. ![](https://terawatttimes.org/content/images/2026/09/Fig-3-4.webp) Figure 3\. The Apprenticeship Layer Under Intelligent Substitution ### 7.2 Social Protection as a Condition of Change Social protection does two jobs at once. It stabilises consumption, and it lowers resistance to technological change, since workers accept displacement more readily when health care, income support, pensions and access to training stay with them after a job ends. Without those protections, resistance to automation is a rational response to concentrated personal risk. The coverage gradient in Section 6 gives the point its scale. In low-income countries, where fewer than one person in ten has effective social protection coverage \[10\], the cost of displacement falls almost entirely on the household, and so does the incentive to resist. Where coverage is broad, the same technology meets a workforce that can afford to move. At the state scale the corresponding instruments are diversified sources of critical inputs, open and interoperable standards, domestic maintenance and integration capability, strategic inventories, and the capacity to reconstruct critical functions through a new technical paradigm. Together they give productive autonomy its operational meaning, sufficient control to operate, repair, replace and negotiate. This pillar fails when income stops during transition. A retraining entitlement that pays nothing while the worker retrains is available on paper and unusable in practice, a nominal alternative that is listed and out of reach. ## 8 Pillar Three: Peaceful Correction Correction mechanisms have to operate before accumulated pressure becomes a crisis. They include public access to information, independent auditing, competition and interoperability rules, judicial and legislative revision, algorithmic impact assessment with rights of explanation and appeal, and international dispute settlement. AI cuts both ways here, and the evidence is concrete. OECD work finds that AI can improve access to benefits, raise service quality and efficiency and free staff time, with common applications in client support, back-office automation and fraud detection. The same work identifies substantial risks, among them the dehumanisation of social protection, problems of data quality and exclusion, opacity, threats to public trust, and information asymmetry and vendor lock-in in procurement \[13\]\[14\]. **Vendor lock-in** is the production-control problem of industry reappearing inside the welfare state. Control over a system can sit with whoever holds its software, models and update rights, whoever operates it day to day, so a government can run its own social protection system and still depend on another party for what makes it run. The case for diversified model sources and open standards in industry applies with equal force to the administrative capability of the state. ### 8.1 Two Documented Failures of Correction Two public inquiries show what happens when automated or mass-processed administration outruns the mechanisms meant to correct it. Australia's Robodebt scheme compared annual income data held by the Australian Taxation Office with the fortnightly income welfare recipients had declared. It converted the annual figure into fortnightly amounts by averaging and raised a debt wherever the two diverged \[15\]. Averaging assumes that income arrives evenly across the year, which for casual, seasonal and part-time workers it rarely does. Recipients who received a notice had to disprove the debt, shifting the onus of proof from the agency to the individual \[15\]. The scheme drew heavy public criticism by the end of 2016 and continued until November 2019, when the government announced that debts would no longer be raised solely on the basis of averaged income, and in 2020 a class action was settled and all debts raised wholly or partly through averaging were reduced to zero \[15\]. The Royal Commission established in August 2022 reported in July 2023\. It found that the scheme produced inaccurate results and did not comply with the income calculation provisions of the Social Security Act, and it called Robodebt a "crude and cruel mechanism, neither fair nor legal" \[15\]. Its recommendations address automated decision-making, data matching, and accountability and capability in the public service \[15\]. In July 2020 the Dutch House of Representatives appointed a parliamentary committee of inquiry into the tax administration's anti-fraud approach to childcare benefits. Its report, *Ongekend onrecht* ("Unprecedented injustice"), presented in December 2020, found that fundamental principles of the rule of law had been violated \[16\]. Errors in applications were quickly treated as fraud, and parents were wrongly branded as intentional fraudsters. The committee placed responsibility on the executing agency, the legislature and the judiciary alike, found the legislation so rigid that it left too little room to do justice to individual situations, and found that the ministry responsible for implementation had run the benefit as a mass process \[16\]. Four features recur across the two cases, and each corresponds to a requirement of this pillar. Scale multiplied error, because a flawed rule applied by automation or mass processing reproduces its flaw in every case it touches, and correction capacity therefore has to grow with the degree of automation in administration. The onus moved to the weaker party, who had to prove innocence against a system holding the data; rights of explanation and appeal exist to keep that onus where the information sits. Responsibility dispersed, most visibly in the Dutch case, where it ran through the agency, the legislature and the courts at once and each could point to the others, which is what the requirement of identifiable responsibility prevents. And correction arrived late. Both schemes were eventually corrected through peaceful means, with debts cancelled and inquiries establishing what had happened, but the correction came years after the harm and through extraordinary instruments. Speed was the variable that failed, which is why periodic review and standing appeal routes matter more than any single inquiry. Correction space, in other words, existed in both countries. What it lacked was the ability to act before accumulated harm became a political crisis, and that is the standard this pillar sets. Correction speed is measured by two intervals in the public record: from the first documented complaint or adverse ruling to the suspension of the practice, and from suspension to the moment affected people receive a remedy. Institutional responsibility must remain identifiable. An agency or company cannot escape accountability by attributing a decision to an algorithm, and people affected by consequential automated decisions need a route to understand, challenge and correct them. The same principle covers emergency restrictions and decoupling measures, which require periodic review and an institution authorised to reassess them if crisis-era controls are to avoid hardening into permanent concentrations of power. Visible calm can mean two things, so correction space is measured through institutions. The observables are whether people can organise, bargain collectively, obtain relevant information, challenge decisions, take part in policymaking and receive a genuine institutional response; a fall in protest under restrictive conditions leaves the state of compression unknown. The same logic holds inside firms and public administrations, where workers who stop reporting problems may be satisfied or may have learned that reporting achieves nothing, and the institutional conditions tell the two apart. External conflict, for its part, reallocates pressure and generally raises total survival risk. It can reorganise domestic coalitions and accelerate industrial mobilisation, and it also destroys infrastructure, disperses skilled populations and eliminates productive assets, which intensifies the original problem. The characteristic failure of this pillar is a set of crisis-era controls that persist without periodic review. A restriction adopted for an emergency and never reassessed becomes a standing concentration of control, outside the correction mechanisms the pillar exists to provide. ## 9 Stable Rights and Adjustable Parameters: A Cross-Pillar Design Principle Shared-output governance faces one central institutional tension. Claims must be stable enough to support decisions about housing, education, family life, consumption, investment and retirement, since constant discretionary change weakens trust and makes long-term planning difficult. The level and composition of sharing must also respond to productivity, prices, demographics, fiscal capacity and technology, because a permanently fixed formula eventually becomes inadequate or unsustainable. A durable design separates the stability of entitlement from the adjustment of parameters. Basic eligibility and protections are established as rights, while benefit levels, contribution rates, dividend formulas and working-time standards change through transparent and reviewable procedures that publish the relevant data, identify adjustment rules, permit participation and provide opportunities for appeal. Applied to a public wealth fund or a social dividend, the principle takes a concrete form. Eligibility is set in statute. The formula linking fund income to payments is published with the data that feed it, and any change follows a scheduled review, with public participation and a route of appeal, taking effect only after notice. Households can then plan around the claim even when its level moves. The principle joins two pillars, giving the first its stability and the third a defined role, since parameters move through correction procedures instead of through discretion. Its absence has a documented cost. In 2016 the governor of Alaska reduced the transfer appropriated for Permanent Fund dividends from an estimated 1.362 billion dollars to 695.65 million dollars, and individual payments fell to 1,022 dollars, about half the amount expected. The state supreme court subsequently confirmed that Permanent Fund income remains subject to the normal appropriation and veto processes \[17\]. When the rent contracted, the stable claim was the first thing adjusted, and the adjustment came through veto and litigation with no designed correction mechanism involved. What was lost was less the level than the predictability. Predictability is one of the conditions under which claims convert into consumption, alongside the share of essential costs paid out of pocket and the status of a claim as a right. Comparative work on welfare regimes has long treated the degree to which entitlement is insulated from market and political fluctuation as a defining institutional variable \[18\]. ## 10 Relation to Existing Governance Accounts Many of the instruments discussed here have their own literatures, and each addresses part of the problem. The social investment approach emphasises capabilities built across the life course through education, training and care \[19\]\[20\]. Predistribution seeks more equal market outcomes before taxes and transfers, through wages, bargaining and market rules \[21\]. Basic income offers an unconditional cash claim \[22\], and universal basic services secure essential needs in kind \[23\]. Flexicurity combines labour-market flexibility with income security and activation \[24\]. Just-transition frameworks pair structural change with social dialogue, social protection and skills for affected workers \[25\]. Algorithmic accountability asks how automated decisions can be explained, audited and contested \[26\]. Public wealth funds give the public a claim on capital returns \[8\]\[9\]. Read against the three levers, each account has a centre of gravity. The table records the primary analytical emphasis of each account in its canonical formulations. Table 2\. Primary Analytical Emphasis of Existing Governance Accounts Across the Three Levers | Account (canonical formulation) | Stable claims | Capability reconstruction | Peaceful correction | | ------------------------------- | --------------------------- | ------------------------- | ---------------------------------- | | Social investment | Secondary | Primary | Not a primary focus | | Predistribution | Primary | Secondary | Secondary | | Basic income | Primary | Not a primary focus | Not a primary focus | | Universal basic services | Primary, on essential costs | Secondary | Not a primary focus | | Flexicurity | Secondary | Primary | Not a primary focus | | Just transition | Secondary | Primary | Secondary, through social dialogue | | Algorithmic accountability | Not a primary focus | Not a primary focus | Primary | | Public wealth funds | Primary, on capital income | Not a primary focus | Depends on governance design | Shared-output governance adds three things to these accounts. It derives the levers from a causal account of compression, so each instrument is justified by the variable it moves: claims and correction as moderators, capability reconstruction as the lever on transmission. It treats the pillars as interdependent, since claims make retraining usable, capability makes an appeal worth having and correction keeps claims stable, so a strong pillar beside a weak one is expected to underperform. And it attaches an observable failure mode to each pillar, which turns a design into something that can be audited. On this reading, an institutional system built around one column is an incomplete architecture. Where institutions cover all three levers, the gap between AI exposure and social outcome narrows. ## 11 Conclusion Artificial intelligence raises what an economy can produce, and institutions decide whether it also widens what people can choose. This paper has specified which institutions matter and how they fail. The domestic scale of compression turns on four speeds. When substitution runs faster than new-task formation, personal capability reconstruction and institutional adjustment for a sustained period, choice narrows even where aggregate employment holds. American and Danish evidence shows exposure and effect moving independently, and the IMF analysis points to the same open variable, the pace at which institutions respond. Three levers reach that variable. Stable claims on output combine labour, capital, public-service, cash and time channels, and they fail when entitlement depends on discretionary appropriation or veto. Capability reconstruction combines income during transition, credentials, practice, entry routes and portable protection with the preservation of the apprenticeship layer, and it fails when income stops during transition. Peaceful correction combines auditing, explanation, appeal and review. It fails when crisis-era controls persist without review, and the Robodebt and Dutch childcare cases show speed to be a critical dimension of it. The principle of stable rights with adjustable parameters joins the first and third levers, and Alaska's 2016 dividend cut shows what happens without it. The argument closes on capability reproduction. The capability-reconstruction pillar reaches training systems, occupational entry routes, supplier retention, maintenance competence, certification capacity and the scheduling of work, which are the carriers through which productive abilities are maintained or lost. Governance therefore acts on the dynamics of compression as well as on its effects. Institutions that keep junior work flowing, keep income steady through transition and keep projects in steady sequence change the rate at which capability erodes and the cost at which it can be rebuilt. An economy organised on these lines can absorb what it produces, because people hold claims on output that convert into demand. It can accept technological change, because the risks of change are pooled, and it can keep international exchange open, because dependence remains negotiable and correction remains available. Artificial intelligence can contribute to that order by spreading knowledge, lowering reconstruction costs where the missing element is knowledge and the complements exist, improving public services and expanding the number of productive actors. Those gains last where a society can purchase what it produces, where people share in the result, and where dependent parties retain realistic ways to change course. ## References \[1\] Gmyrek, Pawel, Janine Berg, Karol Kamiński, Filip Konopczyński, Agnieszka Ładna, Balint Nafradi, Konrad Rosłaniec, and Marek Troszyński. *Generative AI and Jobs: A Refined Global Index of Occupational Exposure.* ILO Working Paper 140\. Geneva: International Labour Organization, 2025\. https://doi.org/10.54394/HETP0387. \[2\] Organisation for Economic Co-operation and Development. *OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market.* Paris: OECD Publishing, 2023\. https://doi.org/10.1787/08785bba-en. \[3\] Lane, Marguerita, Morgan Williams, and Stijn Broecke. "The Impact of AI on the Workplace: Main Findings from the OECD AI Surveys of Employers and Workers." OECD Social, Employment and Migration Working Papers No. 288\. Paris: OECD Publishing, 27 March 2023\. https://doi.org/10.1787/ea0a0fe1-en. \[4\] Brynjolfsson, Erik, Bharat Chandar, and Ruyu Chen. *Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence.* Stanford Digital Economy Lab Working Paper, November 2025\. https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/. \[5\] Humlum, Anders, and Emilie Vestergaard. *Large Language Models, Small Labor Market Effects.* NBER Working Paper No. 33777\. Cambridge, MA: National Bureau of Economic Research, 2025; also University of Chicago Becker Friedman Institute Working Paper No. 2025-56. \[6\] Moder, Isabella, and Til Pommer. "AI and the US Labour Market: Effects on Employment Growth." Box in *ECB Economic Bulletin*, issue 4/2026\. Frankfurt: European Central Bank, 22 June 2026. \[7\] Cazzaniga, Mauro, Florence Jaumotte, Longji Li, Giovanni Melina, Augustus J. Panton, Carlo Pizzinelli, Emma J. Rockall, and Marina Mendes Tavares. *Gen-AI: Artificial Intelligence and the Future of Work.* IMF Staff Discussion Note 2024/001\. Washington, DC: International Monetary Fund, 2024\. https://doi.org/10.5089/9798400262548.006. \[8\] Norwegian Ministry of Finance. *Management Mandate for the Government Pension Fund Global.* Consolidated as at 30 June 2024\. Oslo: Government of Norway. https://www.regjeringen.no/contentassets/9d68c55c272c41e99f0bf45d24397d8c/2024.06.30\_gpfg\_management\_mandate.pdf. \[9\] International Working Group of Sovereign Wealth Funds. *Sovereign Wealth Funds: Generally Accepted Principles and Practices ("Santiago Principles").* Washington, DC, 2008. \[10\] International Labour Organization. *World Social Protection Report 2024–26: Universal Social Protection for Climate Action and a Just Transition.* Geneva: ILO, 2024\. https://doi.org/10.54394/ZMDK5543. \[11\] Brollo, Fernanda, Era Dabla-Norris, Ruud de Mooij, Daniel Garcia-Macia, Tibor Hanappi, Li Liu, and Anh D. M. Nguyen. *Broadening the Gains from Generative AI: The Role of Fiscal Policies.* IMF Staff Discussion Note 2024/002\. Washington, DC: International Monetary Fund, June 2024\. https://doi.org/10.5089/9798400277177.006. \[12\] Folz, Jean-Martin. *La construction de l'EPR de Flamanville: Rapport au Président-Directeur Général d'EDF.* Paris: Ministère de l'Économie et des Finances, October 2019; released 28 October 2019\. https://presse.economie.gouv.fr/1505-rapport-de-jean-martin-folz-la-construction-de-lepr-de-flamanville/. \[13\] Organisation for Economic Co-operation and Development. *AI and the Future of Social Protection in OECD Countries.* OECD Artificial Intelligence Papers No. 42\. Paris: OECD Publishing, 2025\. https://doi.org/10.1787/7b245f7e-en. \[14\] Organisation for Economic Co-operation and Development. *Harnessing Artificial Intelligence in Social Security: Use Cases, Governance and Workforce Readiness.* OECD Digital Government Studies. Paris: OECD Publishing, 2025\. https://doi.org/10.1787/b52405c1-en. \[15\] Royal Commission into the Robodebt Scheme. *Report.* Commonwealth of Australia, 7 July 2023. \[16\] Parlementaire ondervragingscommissie Kinderopvangtoeslag. *Ongekend onrecht* \[Unprecedented injustice\]. Kamerstukken II 2020/21, 35 510, nr. 2\. The Hague: Tweede Kamer der Staten-Generaal, 17 December 2020\. https://www.officielebekendmakingen.nl/kst-35510-2.pdf. \[17\] Alaska Supreme Court. *Wielechowski v. State*, 403 P.3d 1141 (Alaska 2017). https://law.justia.com/cases/alaska/supreme-court/2017/s-16558.html. \[18\] Esping-Andersen, Gøsta. *The Three Worlds of Welfare Capitalism.* Princeton, NJ: Princeton University Press, 1990. \[19\] Morel, Nathalie, Bruno Palier, and Joakim Palme, eds. *Towards a Social Investment Welfare State? Ideas, Policies and Challenges.* Bristol: Policy Press, 2012. \[20\] Hemerijck, Anton, ed. *The Uses of Social Investment.* Oxford: Oxford University Press, 2017. \[21\] Hacker, Jacob S. "The Institutional Foundations of Middle-Class Democracy." In *Priorities for a New Political Economy: Memos to the Left*, 33–38\. London: Policy Network, 2011. \[22\] Van Parijs, Philippe, and Yannick Vanderborght. *Basic Income: A Radical Proposal for a Free Society and a Sane Economy.* Cambridge, MA: Harvard University Press, 2017. \[23\] Coote, Anna, and Andrew Percy. *The Case for Universal Basic Services.* Cambridge: Polity, 2020. \[24\] Wilthagen, Ton, and Frank Tros. "The Concept of 'Flexicurity': A New Approach to Regulating Employment and Labour Markets." *Transfer: European Review of Labour and Research* 10, no. 2 (2004): 166–186. \[25\] International Labour Organization. *Guidelines for a Just Transition towards Environmentally Sustainable Economies and Societies for All.* Geneva: ILO, 2015. \[26\] Diakopoulos, Nicholas. "Accountability in Algorithmic Decision Making." *Communications of the ACM* 59, no. 2 (2016): 56–62.