flowchart LR
B["Nondeployment baseline and affected-population map"] --> S["Staged adoption with pause conditions"]
S --> T["Observed task, role, skill, access, and decision-right changes"]
T --> D["Distribution: wages, prices, ownership, workload, concentration"]
D --> E{"Agency, continuity, remedy, and exit remain reachable?"}
E -- "no" --> R["Compensate, redesign, reduce scope, pause, or withdraw"]
E -- "yes" --> L["Renew bounded deployment lease"]
R --> F["Delayed follow-up and affected-party review"]
L --> F
F --> A["Audit distribution, continuity, agency, and remedy reach"]
A -. "unrepaired burden" .-> R
F -. "material change" .-> B
45 AI Deployment, Transition, Distribution, and Human Agency
45.1 Chapter status
| Field | Value |
|---|---|
| Chapter ID | ai-deployment-transition-distribution-and-human-agency |
| Part | Part II - Planning, Memory, Reasoning, and Execution |
| Status | conceptual |
| Manuscript maturity | v0.3 concept-complete argument-level manuscript |
| Last updated | 2026-08-08 |
| Claim label | Design rationale |
| Evidence level | argument |
| Source loading state | source notes: coherence_exchange, ext_generative_ai_at_work_2025, ext_ilo_genai_jobs_index_2025, ext_oecd_ai_infrastructure_competition_2025; connector/recovery: coherence_exchange |
| Test state | The chapter defines a minimum implementation and falsification plan; no chapter-core promotion follows from prose or source synthesis. |
45.2 Drafting guardrail
This chapter owns the social transition created by deployment: who gains capacity, who loses options, what becomes concentrated, and which remedies remain reachable. It does not infer welfare from productivity or treat the existence of a transition record as evidence that the transition is fair, legitimate, or reversible.
45.3 Human Reading Path
Concrete lens. The simpler baseline reports the aggregate gain of 90. The chapter keeps each cohort and remedy state visible, so the positive total cannot erase the worker cohort’s unresolved burden.
The same AI system can make one worker faster, eliminate another role, create customer dependence, enrich an owner, and make a public service cheaper but harder to contest. These outcomes are different positions inside one transition. They must stay visible before an efficiency story dominates and reversible experiments harden into arrangements that affected people cannot leave.
Technical capability differs from deployment choice, and productivity differs from distribution. A transition packet records the changed task, affected populations, baselines, expected gains, displacement and dependency risks, accessibility, bargaining power, alternatives, notice, authority, appeal, remedy, monitoring, and exit. It asks who receives benefits, who bears costs, whose judgment is displaced, and whether meaningful agency remains.
The lifecycle moves from mapping and participatory design through staged rollout, outcome measurement, redistribution, contest, rollback, and long-horizon review. Deskilling, surveillance, algorithmic management, accessibility failures, vendor lock-in, coercion, benefit concentration, service withdrawal, and ceremonial human oversight remain visible. Deployment is a governed social transformation, not a final API call, and broad benefit must be demonstrated across people and time rather than inferred from average output or lower unit cost.
45.4 Problem
Deployment creates path dependence. Once workflows, skills, data formats, procurement, accessibility adaptations, and customer expectations reorganize around a system, nominal rollback may no longer restore the prior option set. The transition problem therefore includes who pays for adaptation, who can refuse, which alternatives survive, and how delayed harms are repaired after the organization has already changed.
A useful AI deployment changes tasks, roles, skill, discretion, wages, ownership returns, prices, access, concentration, critical-service continuity, and the practical choices available to people and communities. Aggregate productivity cannot reveal who benefits, who loses, or whether transition capacity exists.
Without a transition owner, exposure becomes displacement, adoption becomes benefit, and average productivity hides bargaining power, access, dependency, and distribution. The shared lifecycle method supplies common custody rules; deployment governance uniquely binds affected populations, counterfactuals, delayed outcomes, remedy, and exit.
45.5 Why existing approaches are insufficient
Task benchmarks, exposure indices, adoption counts, productivity averages, job forecasts, and organizational charts describe different levels. Treating any one as a welfare result hides substitution versus complementarity, delayed effects, distribution, bargaining power, ownership, access, deskilling, dependency, and exit.
A composition of Labor OS, Human–AI Organizations, Resource Economics, institutions, and market analysis is the strongest simpler alternative. It wins if it can preserve worker, customer, community, owner, region, service, and supplier outcomes on one denominator while keeping task change, welfare, concentration, agency, and remedy distinct.
The strongest objection is transition bureaucracy: extensive observation and remedy duties risk delaying useful tools, entrenching incumbents, or freezing the baseline, so burden, urgency, proportionality, and non-AI alternatives stay visible.
What this deployment-transition diagram shows: a population- and time-bounded transition lease moves. Productivity never crosses the diagram as a substitute for welfare, distribution, continuity, or agency.
45.6 Core Claim
Reader claim. A deployment can raise total output while harming a worker, removing a practical alternative, and concentrating the gain elsewhere. Productivity is therefore one transition outcome, not a verdict on welfare or agency.
Operational rule. Freeze the nondeployment baseline and every affected cohort before rollout. Track gain, burden, remedy, access, ownership, continuity, refusal, exit, and delayed effects separately; one unremedied harmed cohort blocks a successful-transition claim even when the aggregate is positive.
[ai-deployment-transition-distribution-and-human-agency.core, label: Design rationale, support: argument] Consequential deployment should advance only through a prospective transition contract that binds a counterfactual baseline, affected-person denominator, task-role-skill changes, adoption, substitution and complementarity, compensation and ownership, access and prices, concentration, critical-service continuity, human decision rights, training and redeployment, delayed outcomes, remedy, pause conditions, and residuals; exposure, productivity, adoption, or aggregate gain alone establishes neither job loss, welfare, fairness, human agency, nor a successful transition.
45.7 Mechanism
45.7.1 Worked distribution ledger: positive total, unresolved worker harm
The authored transition dossier contains two cohorts. Cohort 1 receives a gain of 100 with no recorded burden. Cohort 2 receives no gain, bears a burden of 10, and has received no remedy. Both remain in the denominator, so the aggregate is positive 90. An average-only report would call the transition a success and erase the second cohort behind the first.
The transition contract refuses that move. Its finite-list rule preserves every expected cohort and lets one unremedied harmed member block acceptance. Fifty-four single-axis mutations test the surrounding obligations: baseline, worker, customer and community denominators, attrition, task/role/skill separation, compensation and ownership, service continuity, practical refusal, training and redeployment, delayed follow-up, remedy funding, pause, withdrawal, and residual ownership. The numbers are an authored counterexample, not a field estimate, but they make the book’s distribution claim impossible to misread: a positive sum does not cancel an unresolved person.
The transition contract starts with a causal map rather than an adoption plan. It identifies the unit of deployment, the plausible nondeployment path, affected tasks and services, people and institutions exposed directly or indirectly, expected substitution and complementarity, ownership and supplier relationships, and the time horizons on which skill, wages, access, and concentration could change. Uncertain edges remain hypotheses to measure.
Rollout stages are designed as reversible comparisons. Each stage has an eligible population, capacity and support prerequisites, preserved alternative, measurement window, pause threshold, and funded remedy. Adoption, actual use, quality, throughput, workload, error burden, discretion, learning, compensation, prices, accessibility, portability, and service continuity are recorded separately. Hidden labor and knowledge contributions are assigned to the people and organizations that supplied them instead of disappearing into “model productivity.”
Distributional accounting follows gains and losses to workers, customers, owners, suppliers, regions, public services, and future entrants. A positive aggregate does not settle whether the rollout is acceptable. The decision packet asks which groups improved, which were harmed, whether losses were avoidable or compensated, whether bargaining power changed, and whether practical exit remains. Delayed review can reopen the lease when deskilling, dependency, concentration, or service degradation appears after the initial productivity window. This preserves the option to act on late evidence instead of treating launch as an irreversible success declaration.
Contract. Freeze the deployment, nondeployment counterfactual, affected-worker and community denominator, rollout stages, measurement schedule, and pause criteria before adoption.
Admission. Track exposure, actual use, task change, role change, employment, skill, discretion, workload, compensation, ownership returns, prices, access, quality, and distribution separately.
Execution. Map substitution, complementarity, human-AI decision rights, hidden labor, data and expertise contributions, bargaining power, concentration, and realistic exit.
Observation. Fund training, redeployment, income and service continuity, accessibility, contestability, and remedy as deployment costs rather than externalities.
Closure. Use staged rollout and delayed follow-up to narrow, pause, redesign, compensate, or withdraw when transition capacity or observed outcomes fail.
45.8 Concept-completion ledger
45.8.1 Distribution of gains and burdens
Mechanism. Build a distribution ledger before rollout that names workers, customers, owners, suppliers, contractors, data and knowledge contributors, regions, public services, and excluded or future entrants. For each cohort, track quality, time, workload, compensation, prices, access, errors, surveillance, transition cost, and ownership returns over declared horizons. Aggregate benefit is reported beside subgroup outcomes and uncompensated contributions; a positive total cannot close an unresolved burden.
Failure mode. Average productivity can hide wage suppression, shifted error work, inaccessible services, uncompensated expertise, or gains captured by owners and infrastructure providers. Narrow accounting can also double-count benefits or treat ordinary process improvements as AI effects. An overbroad ledger may become costly without changing a decision.
Non-claim. Distributional visibility does not establish a uniquely fair allocation, causal attribution, economy-wide welfare, or entitlement to deploy.
Source grounding. ext_generative_ai_at_work_2025 supplies bounded worker-level productivity and heterogeneous-effect evidence in one customer-support setting. ext_ilo_genai_jobs_index_2025 supplies heterogeneous exposure mapping, not realized outcomes. coherence_exchange motivates value/accounting and contestability as authorial design lineage. None proves broad distributional benefit.
45.8.2 Labor transition from tasks to durable capability
Mechanism. Trace exposure to actual adoption, task change, role redesign, employment, skill acquisition or loss, workload, bargaining power, wages, and delayed mobility as separate transitions. A deployment contract funds training, protected learning time, redeployment, income continuity, and a non-AI comparator that receives equal process-improvement attention. It records whether assistance complements workers, substitutes for tasks, changes hiring, or transfers tacit expertise into firm assets.
Failure mode. Occupational exposure can be mislabeled displacement, a short productivity gain can be called durable upskilling, and survivor bias can omit people who leave. Management pressure can make “voluntary” adoption coercive. Conversely, a study can manufacture a negative result by withholding adequate training from the AI-assisted group.
Non-claim. One task or workplace cannot establish general employment, wage, skill, or macroeconomic effects, and exposure estimates are not forecasts.
Source grounding. ext_ilo_genai_jobs_index_2025 explicitly separates task exposure from realized automation and reports cross-group variation. ext_generative_ai_at_work_2025 reports heterogeneous productivity effects within a specific firm and workflow. Neither establishes long-run labor transition or an ASI-era conclusion.
45.8.3 Human agency and meaningful choice
Mechanism. Measure agency through practical options: the ability to understand a decision, exercise discretion, refuse AI mediation without retaliation, contest an output, reach a human with authority, select an alternative, and recover from error. The record includes power asymmetry, essential-service dependency, surveillance, workload targets, accessibility, and the cost of dissent. Nominal approval counts only when the person retains a usable path to say no.
Failure mode. Ceremonial human review can coexist with quotas, ranking systems, or service designs that make disagreement costly. A checkbox can stand in for consent, while automation removes the skills needed to overrule it. Paternalistic governance can also deny people useful tools in the name of preserving a choice they did not want.
Non-claim. More human clicks do not prove agency, dignity, welfare, informed consent, or better decisions.
Source grounding. coherence_exchange motivates fork, exit, audit, and contestability as speculative governance interfaces. ext_generative_ai_at_work_2025 examines bounded work-experience outcomes but does not validate meaningful-control mechanisms. ext_ilo_genai_jobs_index_2025 identifies heterogeneous exposure without measuring individual agency.
45.8.4 Access, concentration, and bottleneck power
Mechanism. Map control over chips, fabrication, cloud, energy, networks, models, data, app distribution, identity, and contractual interfaces. Record supplier concentration, upstream common dependencies, pricing and capacity terms, vertical integration, switching cost, data egress, portability tests, and whether smaller organizations or regions can obtain a functionally adequate alternative. A deployment lease treats migration capacity and service continuity as first-class costs.
Failure mode. Several branded vendors may rely on the same upstream accelerator or cloud, making apparent competition brittle. Technical interoperability can exist on paper while data, capital, power, talent, or contracts prevent migration. Concentration metrics can also be overread as proof of illegality or as evidence that fragmentation is always safer.
Non-claim. This map does not define a legal market, prove anticompetitive conduct, select an optimal industrial policy, or show that decentralization improves every outcome.
Source grounding. ext_oecd_ai_infrastructure_competition_2025 maps the AI infrastructure supply chain, barriers, vertical relationships, and policy concerns with time- and jurisdiction-specific limits. coherence_exchange adds authorial fork/exit framing. No source adjudicates a particular provider or validates the chapter’s portability gate.
45.8.5 Exit, fork, and contestability
Mechanism. Define effective exit as a rehearsed transition of data, workflows, identities, interfaces, accessibility adaptations, audit history, and essential-service obligations to a viable alternative. Fork rights specify which artifacts and governance state may move; contestability supplies notice, explanation, evidence access, appeal, remedy, and an authority able to change the result. Exit tests include provider withdrawal, price change, policy change, degraded quality, and a disputed automated decision.
Failure mode. A downloadable file can masquerade as portability when proprietary formats, learned workflows, or upstream dependencies remain stranded. Formal appeal can be useless if it is slow, inaccessible, or lacks remedial power. Mandatory portability can also expose sensitive data or undermine legitimate integrity controls.
Non-claim. A successful rehearsal does not establish competitive markets, complete independence, lawful entitlement to every artifact, or freedom from all switching cost.
Source grounding. coherence_exchange supplies fork, exit, audit, and contestability as governance design lineage, not tested institutional economics. ext_oecd_ai_infrastructure_competition_2025 grounds the distinction between technical compatibility and practical substitutability. Neither proves that the proposed exit contract restores agency.
45.8.6 Capability and skill preservation
Mechanism. Identify human and organizational capabilities that must survive deployment: diagnosis, exception handling, domain judgment, maintenance, teaching, emergency operation, and supplier migration. Preserve them through deliberate practice, rotation, fallback drills, documentation, apprenticeship, and periodic unaided performance checks. Skill measures separate short-term output assistance from learning, dependency, and the ability to recover after tool withdrawal.
Failure mode. A system can raise immediate throughput while eroding the judgment needed to detect its errors or operate during outages. Unaided testing can become punitive surveillance, and preserving obsolete tasks can waste effort or deny workers access to genuinely useful augmentation. Weak training can then be blamed on the idea rather than implementation quality.
Non-claim. Preserved benchmark performance does not prove meaningful expertise, worker wellbeing, labor-market mobility, or that every displaced skill should be retained.
Source grounding. ext_generative_ai_at_work_2025 motivates separate analysis of productivity, worker experience, learning, and heterogeneity in one workplace. ext_ilo_genai_jobs_index_2025 motivates task-level exposure rather than job-level simplification. Neither supplies long-run skill-retention evidence or validates these drills.
45.8.7 Monitoring, correction, and reachable remedy
Mechanism. Keep the transition lease open through delayed observation. Independent monitors track adoption and nonadoption cohorts, subgroup outcomes, complaints, service failures, concentration changes, attrition, and late dependency. Threshold breaches trigger funded actions—training, compensation, accessibility repair, scope reduction, pause, supplier migration, or withdrawal—and closure requires evidence that the remedy reached the affected state, not merely that it was announced.
Failure mode. Short follow-up can miss deskilling, wage effects, induced demand, or concentration. Complaint systems can exclude precarious workers and inaccessible users, while management-selected metrics suppress harms. A rigid threshold can also pause beneficial service because of noisy data or a comparator deprived of ordinary improvements.
Non-claim. Monitoring does not itself repair harm, prove causal attribution, or guarantee that every affected person can be found and restored.
Source grounding. ext_generative_ai_at_work_2025 offers a bounded field-study comparator and heterogeneous outcomes, not a general remedy system. ext_ilo_genai_jobs_index_2025 helps define prospective affected cohorts. ext_oecd_ai_infrastructure_competition_2025 motivates monitoring of evolving bottlenecks. No source validates effect-complete remedy here.
45.9 Interfaces
Interfaces exchange disaggregated facts rather than a deployment score. Labor OS reports task and role changes; organizations report decision rights and accountability; economic owners report costs, prices, ownership, and concentration; accessibility and service owners report continuity; institutions return duties and remedies. This chapter joins them around the same affected population and timeline.
Labor OS supplies task-level changes; Human–AI Organizations supplies internal roles; Resource Economics supplies infrastructure cost; institutions supply lawful remedies. This chapter joins their external effects across workers, customers, communities, owners, suppliers, and public services without absorbing any neighbor’s authority.
- Labor OS owns a typed work unit; this chapter owns cumulative task-to-social transition.
- Human-AI Organizations owns internal roles and accountability; transition governance includes workers, customers, communities, and markets.
- Resource Economics owns compute allocation; this chapter owns distribution of deployment benefits and burdens.
- Multi-Agent Dynamics reports concentration and disempowerment; public institutions receive cross-organization effects.
45.10 Invariants
The counterfactual and population cannot be rewritten after favorable groups adopt. Attrition, nonuse, denied access, and people who leave the organization remain in the denominator when relevant. A remedy is not complete when funding is announced; receipt, usability, timing, and restoration of practical options must be observed.
These transition invariants specialize the shared lifecycle method: population denominators persist through delayed follow-up, benefit cannot cancel unremedied harm, and nominal choice does not count as exit when essential services, data, or learned workflows are stranded.
- Exposure, adoption, productivity, task change, job change, wages, welfare, access, ownership, and distribution remain separate claims.
- Aggregate gains never erase harmed subgroups or uncompensated contributors.
- Productivity does not stand in for worker, customer, or community welfare.
- Essential-service access, human discretion, contestability, and exit remain measured.
- Deployment pauses when required transition capacity or remedy is absent.
45.11 Failure modes
Common designs create false confidence through survivor bias, voluntary-adopter bias, novelty effects, management pressure, short follow-up, and a comparator that is starved of ordinary improvement. Another failure calls worker review “human control” while quotas, ranking, surveillance, or loss of alternatives make disagreement costly. Natural evaluation must make these pressures and missing observations visible.
The principal failure family includes exposure-as-displacement overclaim; automation theater; deskilling; dependency; wage suppression; uncompensated knowledge capture; rent concentration; regional disparity; digital exclusion; inaccessible services; surveillance productivity; hidden labor; induced demand; political capture; gradual disempowerment.
Evaluation must preserve heterogeneous effects and delayed outcomes rather than optimize one average. A valid study must be able to detect both useful complementarity and real displacement or dependency, and it must record whether compensation, training, portability, and service continuity actually reached affected people.
45.12 Minimum Viable Implementation
The first study should select a workflow where quality, throughput, worker experience, customer outcomes, and rollback are observable without exposing participants to severe irreversible harm. Baseline operations receive equal process-improvement attention. The deployment budget includes training, accessibility, independent measurement, transition time, portability rehearsal, and compensation for burdens created by the study.
Run a staged natural workflow study with a frozen non-adoption comparator, worker and customer measures, subgroup denominators, delayed follow-up, independent analysis, and real remedy. Measure useful throughput and quality alongside skill, workload, discretion, compensation where appropriate, access, distribution, concentration proxies, exit, and transition cost. Simulation can test accounting, not economy-wide welfare.
The minimum study uses a staged natural workflow with a frozen nonadoption comparator, affected-worker and customer cohorts, subgroup denominators, delayed measurement, independent analysis, and a funded remedy path. It reports productivity beside skill, discretion, workload, access, compensation, concentration, exit, and total transition cost.
45.13 Evidence and falsification program
Argument exit requires a prospective deployment study with an explicit counterfactual, adoption and nonadoption denominators, heterogeneous task and welfare outcomes, supplier and ownership effects, delayed follow-up, practical exit tests, and independently observed remedy. No single workplace or service setting can establish economy-wide welfare or an ASI transition.
45.14 Mature Research Target
The mature layer is a transition observatory and control plane spanning organizations and time. It can represent several plausible counterfactuals, connect task-level changes to roles and services, and track how costs and gains move through ownership and supplier networks. Decision makers can inspect average outcomes and distributions, test portability, and see which groups or regions lack capacity to benefit.
Research combines staged field studies, administrative and service data, qualitative observation, organizational experiments, market structure, and long-horizon follow-up. Strong comparators include process improvement without AI, alternative tools and suppliers, worker-directed adoption, and deployment with different ownership or compensation rules. Outcomes distinguish productivity, quality, learning, workload, wages, prices, access, agency, concentration, continuity, and remedy.
No evidence from one workplace can establish an economy-wide transition. The target instead supports cumulative bounded knowledge: which deployment designs produce durable complementarity, where substitution creates uncompensated harm, when concentration makes exit fictional, and which remedies restore options. The control plane should pause or narrow a rollout when measurement capacity and transition resources are weaker than the consequences at stake.
The mature transition layer lets institutions stage, compare, pause, compensate, redesign, or retire deployments while keeping affected parties and delayed consequences visible. It is successful only when capability gains improve durable options and service quality without hiding concentration, dependency, or unpriced transition burdens.
This is not a current-result claim. Transition-governance support should stay at argument until prospective deployments establish credible counterfactuals, delayed distributional measurement, reachable remedy, service continuity, and exercised exit.
45.15 Formalization hooks
AsiStackProofs.DeploymentTransitionGovernance contains 44 theorem declarations for the chapter’s bounded transition-record envelope. An eight-transition lifecycle preserves deployment and baseline identity, complete affected-person denominators, disaggregated accounting, practical refusal and exit, transition capacity, remedy, expiry, and explicit non-authority over arbitrary run length. A complete dossier reaches only a Project Theseus governed transition study; all 54 admission-axis mutations reject readiness and receive an exact repair or refusal disposition.
The finite cohort model proves append composition, expected-cohort coverage, and that an unremedied cohort blocks acceptance. Its concrete counterexample has positive aggregate gain and unresolved worker harm, so aggregate benefit cannot erase the subgroup obligation inside the model. Two further information-loss constructions show why identical productivity and aggregate-gain signals cannot determine whether a harmed cohort remains, and why identical approval and human-review counts cannot determine whether practical refusal is usable. Deployment, baseline, contract, denominator, observation, remedy, and authority changes invalidate transition receipts. Existing Human-AI Organizations, Readiness Gates, and Evidence States consumers reject missing remedy, failed transition checks, and empirical promotion without a transition study.
Chapter support remains argument. The formalization proves authored record custody and local non-inference only; it does not prove field truth, causal deployment effect, job change, welfare, fairness, meaningful agency, lawful remedy, service continuity, deployment readiness, support, release, or external effect. Project Theseus must run the prospective governed transition campaign with competent nondeployment and ordinary-improvement baselines, complete affected-person denominators, delayed subgroup outcomes, exercised exit, continuity failures, and observed remedy.
45.16 Codex test plan
| Test | Purpose | Status |
|---|---|---|
| Exposure versus realized change | Reject any record that infers displacement or welfare directly from occupational exposure. | planned |
| Distribution denominator | Require worker, customer, owner, supplier, regional, and accessibility outcomes rather than an average-only gain. | planned |
| Practical exit | Test portability, continuity, bargaining power, and remedy under a supplier or policy change. | planned |
| Delayed reversal | Verify that pause, compensation, retraining, or withdrawal reaches the affected state after lagged harm appears. | planned |
Implemented formalization route: lean:ai-deployment-transition-distribution-and-human-agency.admission_boundary covers the bounded transition-record lifecycle, cohort accounting, mutation repair, receipt invalidation, and non-identifiability boundary; it cannot establish causal deployment effects, welfare, fairness, meaningful agency, lawful remedy, service continuity, support, release, or transfer.
45.17 Source crosswalk
| Source ID | Title | Bounded use |
|---|---|---|
coherence_exchange |
The Coherence Exchange | Corben-authored speculative lineage connecting epistemic markets, typed labor, verification, alignment, and resource allocation. It motivates cross-layer distribution questions, but it does not measure employment, wages, welfare, concentration, transition capacity, public legitimacy, or preserved human agency. |
ext_generative_ai_at_work_2025 |
Generative AI at Work | Open peer-reviewed field study of a staggered generative-AI assistant introduction among 5,172 customer-support agents, reporting heterogeneous worker and productivity effects in that setting. It is a bounded deployment comparator and does not establish economy-wide employment, wages, inequality, concentration, long-run skill, or ASI-transition effects. |
ext_ilo_genai_jobs_index_2025 |
Generative AI and Jobs: A Refined Global Index of Occupational Exposure | ILO working paper combining task data, worker surveys, expert deliberation, and model-assisted scoring to estimate occupational exposure across countries and groups. Exposure is not realized automation, displacement, welfare, or a forecast of ASI effects, and the study is not a local reproduction. |
45.17.1 Manifest source assignment reconciliation
These rows keep AI Deployment, Transition, Distribution, and Human Agency’s manifest assignments visible at their recorded review boundary. Passage review does not establish local reproduction, performance, safety, deployment, or support-state movement.
| Source | Intake role | Boundary |
|---|---|---|
ext_oecd_ai_infrastructure_competition_2025 |
Metadata-first comparator: Competition in artificial intelligence infrastructure. OECD analysis of concentration, barriers to entry, vertical integration, and competition across AI infrastructure. It motivates bottleneck and exit analysis but does not adjudicate a specific market, legal violation, or optimal remedy. | No passage-level source claim, local implementation, reproduction, safety, performance, deployment, support-state, or ASI result is established by this reconciliation row. |
45.18 Concentration, bottlenecks, and effective exit
Distribution is shaped not only by wages and prices but by control of compute, cloud platforms, chips, models, app stores, data, and distribution channels. The OECD infrastructure competition analysis motivates a deployment-level concentration record without deciding whether any particular firm or practice violates law.
Every consequential rollout should report supplier concentration, vertical dependencies, switching and data-egress costs, interface portability, access to alternative compute and models, common-mode infrastructure, bargaining power, and who can continue the service if a provider changes price, policy, capability, or availability. “Multiple vendors” is not effective exit when they depend on the same upstream accelerator, cloud, identity system, or proprietary format.
Concentration changes agency. Workers, customers, public agencies, and smaller firms may be formally free to leave while losing learned workflows, data, accessibility adaptations, or essential services. The transition contract therefore treats portability, interoperability, migration rehearsal, service continuity, and a funded exit path as deployment costs. Lower unit price cannot launder an arrangement whose practical exit and contestability disappear.
45.19 Summary
Deployment translates technical capability into changed institutions and life options. A transition contract preserves the nondeployment baseline, affected population, rollout stages, task and role changes, skills, workload, compensation, ownership, prices, access, concentration, continuity, decision rights, and delayed outcomes on one inspectable timeline.
The governing question is not whether an AI tool raises an average metric. It is whether benefits endure, burdens and contributions are visible, affected people retain meaningful alternatives, services remain available, and pause, compensation, retraining, redesign, or withdrawal can actually reach the changed state.
Deployment is a governed social transition, not the last step of a technical pipeline. The transition ledger keeps capability, adoption, task change, productivity, distribution, concentration, service continuity, human decision rights, delayed outcomes, remedy, and exit separate so a locally successful tool cannot silently become an irreversible institution.
45.20 Handoff
Artifact Graphs, Audit Logs, and Replay receives the versioned deployment decision, affected-population denominator, rollout events, observed outcomes, remedies, and unresolved residuals. It preserves that history but cannot decide whether the distribution was fair, the remedy sufficient, the practical exit restored, or the next rollout authorized.