Editor’s note: This essay continues the Synthetic Civilization political economy series. The first essay, “Output Without Income,” argued that AI may preserve production while weakening the wage-based social bargain. The second, “The Market Becomes an Interface,” argued that allocation is moving upstream into systems that determine eligibility before buyers and sellers ever meet. The third, “The Wage Was a Legitimacy Machine,” argued that employment did more than pay people; it explained them. The fourth, “Tenants of Intelligence,” argued that the next class divide is ownership versus dependency inside rented intelligence environments. The fifth, “The Compute Estate,” argued that compute infrastructure is becoming the new ground of political economy: the territory on which synthetic production runs and rent is collected. The sixth, “Capital Without Justification,” asked whether capital can still claim the full surplus generated by systems built on public science, collective data, and inherited civilization. The seventh, “Surplus Humans and the Politics of Containment,” examined what happens when people retain claims to income, standing, recognition, and membership after the productive system has learned to operate with less need for their labor. The eighth, “The Tax State After Labor,” turned from the management of surplus populations to the fiscal architecture that must fund that management, asking what happens when the payroll system that once made citizens legible to public authority begins to thin while the surplus of synthetic production migrates into structures the state can no longer easily see or reach. This ninth essay turns from fiscal capacity to allocative capacity: what happens when the state remains accountable for decisions increasingly produced inside technical systems it does not fully control, understand, or readily replace.
The state outsourced judgment but retained liability.
The modern state justified itself as the institution of final decision.
It would decide who received a pension and who did not. It would decide which businesses could operate, who crossed the border, who received the subsidy, who qualified for the treatment, who was admitted to the public housing waitlist, who got the license, who faced the sanction. Decision authority was the substance of sovereignty. To govern was to allocate: to determine, through visible law and accountable human procedure, who received what from the collective systems that citizens paid for and depended on.
Something fundamental has shifted in that arrangement. Not because states are retreating. States are, by many measures, larger and more present in daily life than at any prior point in modern history. They spend more, regulate more, intervene more, and monitor more than their predecessors did. The apparatus has not shrunk.
What has changed is where the decisions actually happen.
More and more, the acts that determine whether a person receives credit, whether an application is approved, whether a benefit is granted, whether a vendor is eligible, whether a student qualifies, whether a worker is hired, whether a defendant is detained, whether a patient is prioritized are not made by a civil servant sitting across a desk. They are generated by systems: scoring models, eligibility algorithms, risk engines, procurement filters, hiring platforms, diagnostic tools, classification architectures, and compliance layers that process inputs and produce outputs before any human official acts.
The state no longer always decides.
It certifies, audits, supervises, and absorbs the consequences.
That is the allocation state: the political form that emerges when governance functions have migrated into technical systems, but the liability and legitimacy of those functions remain publicly assigned to the state. The allocation state is not the state that decides everything. It is the state that must justify decisions made elsewhere. The threshold is not the mere use of software. It is reached when the state can no longer independently reconstruct or readily override the systems that produce its allocations, yet remains the institution the affected person must ask for an explanation.
This is not merely algorithmic government. Algorithmic government describes a tool. The allocation state describes a position. The tool changes what the state does. The position changes what the state is.
From Fiscal Crisis to Institutional Crisis
The previous essay in this series described the state’s fiscal predicament after labor thins: the payroll system that once made citizens legible to tax authorities is losing coverage as production reorganizes around AI, and states must search for new revenue bases in an economy whose surplus increasingly accumulates in structures they cannot easily reach.
The fiscal crisis is only one layer of what is happening to the state.
The deeper institutional crisis runs underneath it. The state is not only struggling to fund itself from an economy that has moved. It is also losing direct command over its own allocation functions. Fiscal capacity and allocative capacity are related but not identical. A state can in principle solve the revenue problem while the allocation problem deepens, finding new ways to extract from AI-generated surplus while the systems through which it governs become progressively less its own.
The allocation state names that second, quieter transformation.
The State Does Not Disappear. It Changes Position.
The story of AI and government is usually told in one of two registers.
The first is fear: surveillance states, automated repression, algorithmic control, the dystopian merger of state power and machine speed. The second is aspiration: digital government, AI-assisted services, efficient public administration, faster and more accurate decision-making at scale. Both registers share a common premise: the state remains the primary actor in its own governance functions. One version is empowered by AI; the other is corrupted by it. In both, the state is still the entity doing the governing.
Working through what is actually happening requires setting that premise aside.
What is underway is neither the empowering nor the enslaving of the state. It is the repositioning of the state from the seat of allocation to the frame around it. The state provides the mandate, absorbs the legitimacy demands, certifies the process, manages the appeals, backstops the failures, and defends the outcome in court. But the outcome was produced elsewhere.
This is the key structural distinction: the visible sovereign and the hidden execution layer have separated. Citizens see the state and hold it accountable. The judgment they are holding it accountable for was generated somewhere the state does not fully control.
This is a change in constitutional position, not only in administrative technique. Constitutions describe the state as the repository of public authority: the entity that taxes, spends, licenses, restricts, and distributes in the name of the public. When the systems making those decisions are private, technically complex, unelected, and difficult to audit, the constitutional description becomes progressively less accurate as a guide to where power actually sits.
The state retains the legal form of sovereignty while exercising less of its substance.
States have always relied on private actors for certain functions: road construction, military logistics, outsourced IT. The novelty lies in the depth and character of the current migration. What is moving upstream is not infrastructure, which the state can rebuild, or logistics, which the state can redirect. What is moving upstream is judgment, the cognitive core of the allocation function that was supposed to be the distinctive contribution of accountable human governance.
When the judgment migrates, what remains?
The Judgment-Liability Split
Every allocation decision has two parts: the judgment that produces the outcome, and the liability that attaches to the outcome. Liability here is not only the narrow legal kind. It is the broader answerability that attaches to the state regardless of where the judgment was made: legal exposure, political accountability, and the public duty to explain.
For most of the modern state’s history, both parts lived in the same institution. A human official exercised judgment and was answerable for it. The civil servant who denied a claim, approved a license, or sanctioned a business could be questioned, audited, and overruled. The decision had an author. That authorship was the mechanism through which democratic governance claimed to be distinguishable from arbitrary power.
Algorithmic systems separate these two parts.
The judgment moves into the system. The welfare algorithm determines the benefit. The predictive tool informs the detention decision. The hiring filter removes the résumé before any human sees it. The procurement system excludes the vendor before any official reviews the bid. Each system encodes a set of choices, priorities, and weightings that determine the output. Those choices were made by someone: a private vendor, a development team, a training dataset, a procurement contract. But the choices are not visible to the applicant. They are often not fully visible to the agency deploying the system. They were made before the specific case being decided was ever presented.
The liability, however, remains with the state.
When the algorithm produces a wrongful denial, citizens do not sue the vendor’s training team. They sue the agency. When the risk model generates a discriminatory pattern, the political accountability runs to the minister, not the model developer. When the automated welfare system removes a benefit erroneously, the grievance is directed at the state, because the state is the institution that promised to administer the system fairly and in accordance with law.
This is the core dysfunction of the allocation state: legal sovereignty and political accountability for decisions the state neither fully made nor can fully explain. The gap between accountability and authorship grows as technical systems become more complex, more entangled with private vendors, and more difficult to audit from the outside.
A human official who makes a bad decision can be questioned, and can explain her reasoning. An algorithm that produces a bad output may have no single reasoning that can be surfaced. It was produced by a model trained on historical data, with weights adjusted through optimization processes, combined with real-time inputs, and processed through a pipeline that no one in the agency fully designed or currently understands.
The decision has an output. It does not always have an accessible author.
The Legitimacy Wrapper
Faced with this gap, the state adapts through legitimacy production.
It generates process: the appearance of accountable governance around decisions that were actually made by systems operating faster and more opaquely than the governance overlay can follow. The state builds audit requirements, algorithmic impact assessments, explainability mandates, human-in-the-loop rules, appeal procedures, vendor certifications, procurement standards, and regulatory frameworks.
None of these instruments is fake. Audit requirements produce at least partial visibility. Appeal procedures allow some corrections. Human review catches some errors. Certification processes screen some bad systems.
But they do not restore what was lost.
The original condition was that a human exercised the judgment, could explain it, and could be held responsible for it. The post-algorithmic condition is that a human reviews or approves a judgment produced elsewhere, under time and volume pressures that often make review nominal rather than substantive, and then holds nominal accountability for a decision the human did not author. The form of accountability remains. The substance has migrated.
What the state is constructing, of necessity, is a legitimacy wrapper: a set of processes and representations that allow the system to maintain the appearance of accountable public allocation even after the effective authority over many allocation decisions has moved into private technical infrastructure.
This is adaptation under constraint, not cynicism. No modern state can refuse algorithmic systems and return to purely manual administration. The volume, complexity, and speed of modern governance make that reversion impossible. The real choice is between algorithmic governance with more or less honest legitimacy architecture around it.
The risk is a subtler one. A state that becomes skilled at wrapping systems in the language of accountability, without restoring its substance, may come to mistake the wrapper for the thing it is wrapping. It may mistake compliance documentation for actual oversight. It may mistake algorithmic audits for democratic accountability. It may mistake the presence of an appeal mechanism for the genuine capacity of citizens to contest decisions made by systems they cannot see or understand.
The wrapper is not the house. The state that settles for the performance of accountability, rather than working to reconstruct its substance, stops building.
Judgment Migrates. Liability Stays.
Three sectors already demonstrate the pattern at scale, each revealing a different dimension of the same structural condition.
Welfare and benefit administration. The Netherlands built a System Risk Indication system called SyRI, designed to identify welfare fraud risk by combining personal data drawn from multiple government sources. In 2020 the District Court of The Hague ruled it unlawful, finding it violated the right to private life under Article 8 of the European Convention on Human Rights: the legal framework was insufficiently transparent and verifiable, and the data processing was disproportionate to the aims pursued. [1] SyRI shows that the split can open even without a private vendor. The state built the system, authorized it, and remained answerable for it. Yet the person it flagged still faced a classification she could not see or meaningfully test, produced by a process no official reconstructed for her. When the system failed the legal test, the state dismantled it. The people it had wrongly flagged had no recourse against the process that produced them.
Criminal justice. Predictive risk assessment tools are used across multiple U.S. states to inform bail, sentencing, and parole decisions. COMPAS, a proprietary tool produced by a private vendor, is the most documented example. A 2016 ProPublica analysis found that COMPAS scores were twice as likely to incorrectly flag Black defendants as high risk compared to white defendants who did not go on to reoffend. The vendor disputed the methodology. Courts disagreed on whether defendants had a right to inspect the underlying algorithm. [2] Throughout the legal argument, the accountability for the sentence ran through the judge. The state signed the sentence. The system determined the input. Whether the defendant could inspect the model at all was itself contested in court.
Healthcare prioritization. In 2019, researchers found that a widely used commercial algorithm allocating healthcare resources showed significant racial bias, providing lower risk scores to Black patients who were comparably ill to white patients, resulting in fewer Black patients being referred for additional care. The source of the bias was structural: the algorithm used healthcare spending as a proxy for healthcare need, and because Black patients had historically received less care due to reduced access, the proxy systematically underestimated the severity of their conditions. [3] The vendor adjusted the algorithm after the study was published. The hospitals that had deployed it remained accountable for the outcomes they had produced in the interim.
Taken individually, each case might appear as a correctable implementation failure. Taken together, they reveal the structural condition the split predicts: in each case the judgment was generated in a system its subjects could not inspect, whether the vendor was private or the state itself, while the mandate and the fallout stayed public.
The judgment-liability split is not a future concern. It is already the operating condition of government across every advanced economy. [4]
The Efficiency Is Private. The Failure Management Is Public.
The allocation state does not only wrap private systems in legitimacy. It also backstops them when they fail.
When the state deploys algorithmic systems that later produce systematic errors, it cannot disown the outcomes.
The efficiency gains from automation accrue to whoever deployed the system: hiring costs fall for the firm, underwriting margins improve for the lender, fraud-detection payroll falls for the agency in the short run. The costs of failure flow disproportionately through the state: remediation programs, litigation exposure, enforcement actions, and political damage arrive at the public institution even when the system that generated the problem was built and sold by a private vendor.
The efficiency is private. The failure management is public.
There is a deeper version of this asymmetry. Where the systems performing the judgment are themselves rented, and increasingly they are, they are intelligence environments the state leases rather than owns, running on compute infrastructure it does not control and cannot easily replicate. In those cases the state is not only outsourcing a task. It has become a tenant of its own allocation functions, paying for positional access to capacity held by someone else. When the lease is where the judgment lives, the efficiency accrues to whoever owns the estate, and the legitimacy burden stays with whoever signed the mandate.
That asymmetry is the structural consequence of deploying privately owned allocation systems in public functions without retaining sufficient operational control to prevent failures rather than only remediate them afterward.
The Allocation State Is Not Omnipotent
A possible misreading of this argument would treat the allocation state as an updated version of the total administrative state: more pervasive than any Weberian bureaucracy, delegating superficially while retaining ultimate control. That reading inverts the actual condition.
The allocation state is frequently dependent on the systems it nominally governs.
This is what distinguishes it from both the classical welfare state, which administered its own allocation functions directly, and the authoritarian surveillance state, which builds or captures technical systems under state authority. The allocation state depends on private vendors, foreign-owned infrastructure, proprietary models, and technical systems it lacks the internal capacity to fully evaluate, replace, or override.
A government agency that has integrated an AI hiring tool cannot easily remove it when its staffing capacity has been reduced on the assumption of automation assistance and its institutional knowledge of alternative approaches has atrophied.
Each adoption of an external system appears rational at the moment: faster processing, lower costs, better accuracy than available alternatives. Each adoption also increases the agency’s dependence on systems whose parameters it cannot set, whose training data it does not own, and whose replacement it cannot easily afford. The dependency compounds across time. Each new integration makes the next exit more expensive.
The allocation state in the Global South illustrates the extended form of this condition. Across large parts of sub-Saharan Africa, South Asia, and Latin America, states outsource allocation to multilateral donor systems, NGO compliance architectures, IMF conditionality frameworks, and World Bank project management requirements. Decisions about what healthcare infrastructure gets funded, which populations qualify for emergency food distribution, which governance reforms unlock budget support are made by institutions not accountable to the citizens they affect. The state signs the agreements, implements the programs, and absorbs the political consequences. The allocation logic was written elsewhere.
Advanced economies are importing a variant of this condition through vendor dependency rather than donor dependency. The structure is the same. The judgment originates outside the state. The liability stays inside it.
Access Is Allocated Before the State Arrives
There is a second dimension of the allocation state less visible in the domain of government services but no less consequential.
The same logic that repositioned the state relative to its own public systems has also produced a private allocation regime that shapes life chances before the state’s systems come into play at all. An earlier essay in this series argued that ranking is itself allocation, and that the decisive act moves upstream of the market: hiring platforms filter résumés, scoring systems decide which applications advance, and a person is sorted long before any public institution is involved. That argument stands. What matters here is what it does to the state.
At no point in that chain is there a publicly accountable allocation system. Each step is private, automated, and governed by the internal policies of whoever built it.
By the time a person reaches the domain where public allocation would apply, welfare benefits, public housing, healthcare coverage, educational access, the private allocation system has already substantially determined the material context in which the public system will find her.
The state allocates at the residual layer. The upstream allocation has already happened.
This is why the allocation state cannot be understood only as a story about government deploying AI. It is about the relationship between private algorithmic allocation systems and public ones, and about which is doing the more consequential work. Public remediation inherits the shape of private allocation. The allocation state manages the residual. It does not design the field.
Sovereignty Becomes Override Capacity
Sovereignty, in the classical understanding, is the capacity for final decision. The sovereign is whoever can say “this is what happens” and make it so.
In the allocation state, that capacity is fragmenting across a chain of actors that no single institution controls. The private vendor built the model. The state deployed it. The state does not fully understand it. The vendor cannot be compelled to explain it completely. No court can inspect it comprehensively. No legislator designed the parameters that produce its outputs. The final decision is distributed across technical choices, contractual relationships, training procedures, and deployment configurations assembled over years by different organizations with different interests.
This is not the end of sovereignty. It is its disaggregation into a form that existing constitutional vocabulary does not yet adequately describe.
Recovering effective sovereignty over the allocation state requires not merely the legal authority to regulate, but four operational capacities that most states currently lack.
Interpretive capacity: the state must know, in substantive rather than formal terms, what the system is doing and why, which requires genuine technical expertise inside government rather than the ability to commission audits no official can evaluate.
Override capacity: the state must be able to stop or modify the system without paralyzing the functions that depend on it, which requires maintaining institutional alternatives even while delegating execution.
Replacement capacity: the state must preserve the ability to switch vendors or rebuild systems before crisis arrives, not after, which means treating that optionality as a governance asset rather than an unnecessary redundancy.
Design capacity: the state must shape the architecture before procurement hardens dependency, which means participating in defining the systems it deploys rather than accepting products built for commercial markets and adapting them to public purposes.
These are the operational definition of what governing means when judgment has migrated into technical infrastructure. Most states, currently, do not have all four. That gap is not a failure of political will. It is the consequence of a transition that moved faster than institutional adaptation could follow, deploying systems whose accumulating dependencies were not visible until they had already formed.
Rights and Scores
Beyond shifting where decisions happen, the allocation state changes the language in which allocation is expressed.
For most of the modern state’s history, allocation was described in legal and moral vocabulary. A person was entitled to a benefit, eligible under a statute, subject to a duty, protected by a right. That language was normative. It organized the relationship between the individual and the state in terms that could be contested, interpreted by courts, and revised through democratic processes.
Algorithmic systems operate in a different vocabulary. Risk scores. Eligibility indices. Fraud probability thresholds. Hiring fit ratings. Benefit tier assignments. Compliance pass/fail determinations. This vocabulary is quantitative, probabilistic, and optimized for operational efficiency. It does not describe the individual. It classifies the individual against a population model. The person is not a rights-holder with specific entitlements. She is a data point whose classification positions her in a distribution.
The legal vocabulary asks: does this person have a right? The algorithmic vocabulary asks: what is this person’s score?
These are not different ways of saying the same thing. Algorithmic allocation does not merely automate legal judgment. It changes the grammar of public obligation from rights to scores. A person can have a legal right and still receive a low score. A person can fail an eligibility index and still be legally entitled. The two systems can produce contradictory outcomes, and when they do, the question of which governs is increasingly the central question of administrative law, civil rights enforcement, and democratic legitimacy.
The allocation state is not merely a state that uses machines to administer law. It is a state in which the tension between normative and operational vocabularies has become a structural feature of governance, not a temporary technical problem that better systems will eventually resolve.
The systems will improve. The tension will remain.
The Allocation State and What Came Before
The preceding essays traced how the AI economy reorganizes the relationship between production and distribution, between labor and claim, between market access and eligibility, between capital and the justification for its returns, between surplus populations and the institutions managing them, and between fiscal states and the assets they can reach.
The allocation state is where those transformations converge at the institutional surface of governance.
Each transformation leaves a remainder no private actor will hold: a benefit to administer, a system to certify, a dependency to define, a political consequence to absorb, a population to stabilize, a revenue base to rebuild.
All of these functions converge on the state, not because the state is the most powerful actor in the AI economy (it often is not), but because the state is the institution society has assigned accountability for outcomes that no other institution has been assigned accountability for.
The allocation state is the form this assignment takes when the state lacks the operational capacity to match its assigned accountability. It is the state held responsible for outcomes it cannot fully govern, legitimizing systems it does not fully understand, managing populations whose claims it cannot fully honor, taxing a surplus it cannot fully reach, and providing a legitimacy wrapper around allocation architectures designed by actors answering to different principals than the citizens living inside the outcomes they produce.
The state has not disappeared. It has been repositioned.
The state does not control the system.
It is what the system calls when it needs to be legitimate.
Notes
[1] Rechtbank Den Haag [District Court of The Hague], NJCM c.s. v. De Staat der Nederlanden, ECLI:NL:RBDHA:2020:865, February 5, 2020. The court ruled that SyRI violated Article 8 of the European Convention on Human Rights, finding the legal basis insufficiently clear and the data processing disproportionate to the aims pursued. The Dutch government subsequently dismantled the system. https://uitspraken.rechtspraak.nl/details?id=ECLI:NL:RBDHA:2020:865
[2] Julia Angwin, Jeff Larson, Surya Mattu, and Lauren Kirchner, “Machine Bias,” ProPublica, May 23, 2016. The analysis found that COMPAS recidivism risk scores produced an asymmetric error structure: Black defendants who did not reoffend were significantly more likely to be classified as high risk than white defendants who did not reoffend. The vendor disputed the statistical methodology. Courts disagreed on whether defendants had a right to inspect the underlying algorithm. https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
[3] Ziad Obermeyer, Brian Powers, Christine Vogeli, and Sendhil Mullainathan, “Dissecting Racial Bias in an Algorithm Used to Manage the Health of Populations,” Science 366, no. 6464 (2019): 447-453. The algorithm used healthcare spending as a proxy for healthcare need. Because Black patients historically received less care due to reduced access rather than lower need, the proxy systematically underestimated the severity of their conditions. https://www.science.org/doi/10.1126/science.aax2342
[4] Virginia Eubanks, Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor (St. Martin’s Press, 2018). Provides the most sustained empirical account of the judgment-liability split operating across welfare, child protective services, and public housing: private or semi-private algorithmic systems producing outcomes for which public institutions bear accountability.
excellent