Beyond Doom, Utopia, and Replacement

I. What This Paper Is Actually Arguing

The argument is not that AI discourse is naive, that optimists are wrong, or that pessimists are right. It is that all three dominant narratives about artificial intelligence share a common analytical blind spot — they treat human beings as the objects of AI's impact rather than as agents with varying capacity to shape the conditions under which AI systems are developed, deployed, and governed. That framing error has concrete governance consequences: it directs institutional attention toward outcomes (employment levels, AI capability benchmarks, extinction probability estimates) while systematically neglecting the conditions that determine whether those outcomes, whatever they are, will be experienced by people with or without meaningful recourse.

This paper uses "human agency" as a specific term, not a general one. It is defined operationally in Section VI across five dimensions — epistemic, economic, political, institutional, and temporal. A governance framework that maximizes aggregate employment without preserving epistemic access preserves some dimensions of agency while destroying others. A framework that prevents catastrophic AI risks without building local institutional capacity may protect the species while leaving specific communities without recourse. The five-dimensional definition is designed to prevent precisely this kind of partial optimization from being mistaken for adequate governance.

Underlying all of these questions is a variable that technological and economic analysis consistently underweights: legitimacy. Societies do not become unstable merely because technology advances, nor because jobs change, nor because institutions struggle to adapt. They become unstable when a sufficiently large portion of the population concludes that the future no longer contains a meaningful place for them. The resulting crisis is not primarily economic. It is civic — the erosion of belief that existing institutions can respond to legitimate concerns through legitimate processes. Artificial intelligence therefore presents a governance challenge that extends beyond productivity, employment, safety, or capability. It presents a legitimacy challenge. The preservation of human agency is ultimately an attempt to preserve legitimacy itself.

The paper's relationship to the Foundation's existing work: the EM Foundation's Transitional AGI Governance paper argues for utility-first deployment as a public trust strategy. The Trust Infrastructure essay argues that governance is necessary regardless of how the personhood debate resolves. This paper occupies the distributional and social space between those two arguments — asking not whether governance is needed (it is) or what it should achieve (public trust), but who bears the costs when governance is inadequate, and what the institutional prerequisites for adequate governance actually are.

II. The Three Narratives — A Taxonomy, Not a Dismissal

The three narratives are presented here as analytical categories, not as strawmen to be defeated. Each has serious intellectual defenders. Each has generated real research programs. Each captures genuine features of the AI landscape that any adequate governance framework must address. The critique that follows is not that the narratives are wrong — it is that each is incomplete in a way that has specific governance costs.

II-A. The Concentration Variant

A fourth concern appears with sufficient frequency and analytical weight in AI discourse to warrant dedicated treatment, distinct from the three primary narratives. It is not the claim that AI replaces everyone, saves everyone, or destroys civilization. It is the claim that AI concentrates power — and that this concentration is itself the governance failure, independent of whether the three narratives' ultimate predictions prove accurate.

Market concentration in AI capability development is well-documented. As of this writing, frontier AI development is concentrated among fewer than ten organizations globally, the majority private firms in two national jurisdictions.¹² The capital and data requirements for frontier development create structural barriers to entry that reproduce this concentration regardless of the pace of capability improvement. Compute concentration is perhaps even more fundamental: advanced AI training and inference require specialized hardware produced by an even smaller number of manufacturers, creating supply chain dependencies that translate directly into geopolitical and commercial leverage.

Data concentration interacts with market and compute concentration in ways that amplify each. The behavioral data, text corpora, and labeled datasets required for frontier AI development are disproportionately controlled by the same organizations that dominate model development — partly through historical first-mover advantage, partly through structural arrangements (data licensing, platform terms of service, acquisition) that systematically consolidate data ownership. This concentration is not merely a commercial fact; it is a governance fact, because the entities best positioned to shape AI development are also the entities with the largest stakes in resisting governance that would constrain them.

Information asymmetry compounds all three forms of concentration. The organizations developing and deploying frontier AI systems possess substantially more information about those systems' capabilities, limitations, failure modes, and societal effects than the regulatory, academic, journalistic, and civil society institutions attempting to assess them. This asymmetry is not incidental; it is structurally reproduced by the pace of development, the commercial sensitivity of capability information, and the technical expertise required to evaluate AI systems independently. The EM Foundation's Corroboration Standard and Assessment Charter are both responses to this asymmetry — attempting to build the institutional capacity for independent evaluation that market structure does not spontaneously produce.

The concentration variant cuts across all three primary narratives. The replacement narrative's distributional question — who captures the productivity gains? — cannot be answered without analyzing the concentration of AI ownership and benefit-capture. The utopia narrative's translation problem — between technical capability and realized social benefit — cannot be solved without addressing the governance structures that determine whether benefits flow broadly or narrowly. The doom narrative's power concentration concern is precisely the concentration variant, though it is sometimes framed in terms of misaligned AI systems rather than concentrations of human and institutional power enabled by AI. In this respect, the concentration variant provides the structural analysis that the three primary narratives approach only partially.

The EM Foundation's Anti-Capture Principle — which requires that the Foundation's assessment systems resist capture from any direction, including industry, donors, governments, political movements, reviewers, providers, and the Foundation's own institutional preferences — is a direct response to the concentration variant. So is the Trust Ledger design, which is explicitly constructed to prevent any single actor from controlling the assessment record. The Assessment Charter's Article IV devotes its governance architecture to ensuring that the concentration dynamics visible in AI development do not reproduce themselves in AI governance. The concentration variant is therefore not a fourth narrative to be evaluated alongside the three primary ones; it is the structural substrate that governance must address if any of the three narratives' stated goals are to be achievable.^26,27

III. What Each Narrative Gets Right — And Why Getting It Partially Right Matters

The Replacement Narrative's Empirical Foundations

The empirical basis for the replacement narrative is substantial and should not be minimized. Frey and Osborne's 2013 analysis estimated approximately 47% of U.S. occupational employment was at high risk of automation within two decades — a figure disputed on methodological grounds but pointing toward genuine structural shifts in which cognitive tasks previously assumed immune to automation became technically tractable.⁴ Acemoglu and Restrepo's subsequent empirical work found that each industrial robot introduced per thousand workers reduced employment by 0.2% and wages by 0.42% in affected commuting zones — documenting that displacement effects are real, locally concentrated, and not automatically offset by job creation elsewhere.⁵

The historical record reinforces the concern. The mechanization of textile production in eighteenth-century England produced structural unemployment for skilled handloom weavers for decades before new employment categories absorbed them and their descendants.¹ The collapse of manufacturing employment in the American Midwest between 1980 and 2010 — driven by both automation and trade — produced lasting increases in mortality, decreased civic participation, and sustained erosion of institutional trust in affected communities.⁶ These are not historical curiosities; they are the calibration data for what inadequate transition governance produces.

The Utopia Narrative's Empirical Foundations

The optimistic narrative also rests on substantive evidence that governance frameworks must take seriously. Brynjolfsson and McAfee document consistent historical patterns in which technologies that appear to eliminate occupations ultimately create new categories of work whose prior nonexistence made them invisible from the pre-transition vantage point.⁷ Goldin and Katz's analysis of the "race between education and technology" found that technological progress consistently created demand for higher-skill workers — a pattern that has sustained per capita income growth alongside substantial occupational transformation throughout the twentieth century.⁸

In domain-specific AI research, genuine capability advances are documented with welfare implications: protein structure prediction at accuracy levels that have measurably accelerated drug discovery;⁹ medical image analysis at specialist-level performance in targeted diagnostic categories;¹⁰ legal document review at speeds that could reduce access-to-justice barriers for individuals priced out of professional legal services.¹¹ A governance framework that dismisses these benefits because it focuses exclusively on displacement costs is as incomplete as one that ignores the displacement costs because it focuses exclusively on benefits.

The Doom Narrative's Empirical Foundations

The risk-focused narrative identifies genuine vulnerabilities that the other two narratives consistently underweight. AI development is currently concentrated among a small number of well-resourced private firms operating primarily in two national jurisdictions, with minimal independent oversight of safety or societal impact.¹² The pace at which capability is advancing creates conditions in which governance institutions — which historically adapt on decade-long timescales — may not keep pace with systems they need to govern.¹³

Misuse is documented rather than hypothetical. AI-generated disinformation has been deployed in electoral contexts. AI systems used for credit scoring, hiring, and criminal risk assessment exhibit systematic disparities across demographic groups.¹⁴ AI-enabled surveillance has been used against journalists, activists, and minority populations in ways technically impossible a decade ago. The doom narrative is correct that these risks are structural rather than incidental — they emerge from the concentration and opacity of AI development, not from the malice of individual actors.

IV. Where Each Narrative's Partial Truth Becomes a Complete Failure

The Replacement Narrative's Characteristic Errors

The replacement narrative makes two recurring errors. The first is treating task automation as equivalent to job elimination without accounting for the economic mechanisms through which new tasks emerge. Automation reduces the cost of producing goods and services; through income and substitution effects, this generates demand for other goods and services, some requiring human labor the automation does not affect.¹⁵ The net employment effect of any automation technology is therefore not determinable from the technology alone — it requires analysis of the full economic context and the institutional response.

The second and more consequential error is aggregating highly heterogeneous effects across geographies, industries, and demographic groups in ways that obscure the distributional question that matters most: who bears the transition costs, and who captures the transition gains. The historical record consistently shows that transitions impose disproportionate costs on specific communities while generating diffuse, slowly realized gains.¹⁶ A narrative focused on aggregate employment outcomes risks treating locally devastating costs as acceptable because the aggregate looks favorable — a form of moral averaging that fails the people most affected and is precisely what the Foundation's five-dimensional agency framework is designed to prevent.

The Utopia Narrative's Characteristic Errors

The optimistic narrative makes two errors that mirror the replacement narrative's. The first is historical selectivity: the transitions cited — electrification, computing, the internet — all involved transitions in which the full benefits took decades to materialize, required substantial public investment in complementary infrastructure, and generated distributional conflict along the way.¹⁷ The productivity paradox described by David — electrification initially reduced measured productivity before eventually transforming it — is not an argument against optimism; it is an argument for patience and for the complementary institutional investments that make technology transitions beneficial rather than merely disruptive.²

The deeper error is conflating technical capability with realized social benefit without accounting for the institutional conditions required to make the translation. Medical AI that performs at specialist level on curated research datasets does not automatically translate into better population health outcomes when deployed through healthcare systems with existing access disparities, liability structures, and workflow constraints.¹⁸ The translation problem — between technical capability and realized social benefit — requires governance institutions that the utopia narrative systematically underweights, because those institutions exist outside the capability research paradigm that generates the narrative's evidence base.

The Doom Narrative's Characteristic Errors

The long-run existential variant of the doom narrative faces a methodological challenge that its proponents acknowledge: the most dramatic predictions are not falsifiable on any short-term timescale, making it difficult to distinguish genuine risk analysis from motivated worst-case reasoning.¹⁹ More practically, focusing institutional attention and scarce governance resources on long-run speculative scenarios risks displacing attention from near-term harms — documented bias, surveillance misuse, information environment degradation — for which governance interventions are both possible and urgently needed.

The power concentration variant, which is better empirically grounded, sometimes slides from the accurate observation that AI development is highly concentrated to the conclusion that deconcentration is the primary governance solution. Deconcentration without accountability does not solve the governance problem; it redistributes it. A world with many small AI developers, none accountable for the systems they deploy, is not meaningfully better governed than a world with a few large ones. The structural problem is the absence of accountable institutions — not the distribution of development activity alone.

V. What Industrial History Actually Shows About Technology Transitions

Five historical transitions are particularly relevant to the AI governance problem: textile mechanization (1760–1850); electrification of American industry (1880–1930); agricultural mechanization (1920–1970); computerization of clerical and manufacturing work (1960–2000); and the internet and platform economy (1995–present). The analysis of these transitions yields three findings that the dominant AI narratives systematically underweight.

The three cross-cutting findings: First, technology transitions impose locally concentrated costs and globally diffuse benefits — the commuting zone analysis in the automation literature is the contemporary empirical expression of this pattern documented across all five historical cases.⁵ Second, transition durations are measured in decades, not years — the aggregate benefits typically materialize only after the worst transition costs have been borne by specific communities who will not share proportionately in the eventual gains. Third, the governance response — its presence, quality, and targeting — is the primary determinant of whether a technology transition produces durable increases in human welfare or durable increases in inequality and social fragmentation.

The AI transition exhibits a feature that makes the historical precedents partly but not entirely transferable: the speed of capability improvement. The five historical transitions above all unfolded over timescales that allowed institutional adaptation — however inadequate that adaptation was — to proceed in parallel with technological diffusion. The current AI transition may be advancing faster than any previous general-purpose technology. If so, the historical lesson about governance becoming structurally harder as capability concentrates becomes more urgent, not less relevant.

VI. Human Agency — An Operational Definition for Governance Purposes

The three narratives share an analytical frame in which human beings are the passive objects of AI's impact. The replacement narrative asks which workers will be displaced. The utopia narrative asks which benefits will be generated. The doom narrative asks what risks will materialize. None of these framings asks what conditions are necessary for people — individually and collectively — to retain meaningful choice about how AI systems are developed, deployed, and constrained. That is the human agency question.

Human agency, as used in this paper, is not a general philosophical concept. It is defined operationally across five dimensions. The operational definition is designed to be measurable — not measured yet, but measurable in principle, which is what makes the Human Agency Index proposal in Section XI tractable rather than merely aspirational.

A critical implication of the five-dimensional structure: agency can be preserved on some dimensions while being eroded on others, and governance frameworks that optimize for one dimension may actively undermine others. A society with high epistemic access to AI systems but no institutional recourse when those systems cause harm has preserved Dimension 1 while destroying Dimension 4. A governance framework that prevents catastrophic AI risks (protecting the temporal dimension) through regulatory structures that exclude affected communities from governance participation (eroding Dimension 3) may protect the species while failing specific populations. The five dimensions must be tracked independently.

VI-A. Modes of Agency Loss

Human agency is not diminished in a single uniform way. The five-dimensional framework captures what is at stake; this section identifies the mechanisms through which each dimension is threatened. Understanding these mechanisms is necessary for governance responses that are specific rather than generic — that target the actual pathway through which agency is being eroded rather than the category of erosion alone. Five distinct modes of agency loss are identified, each with a specific mechanism, documented AI examples, and distinct governance implications. They can occur independently or in combination; the compounding cases are typically the most governance-resistant.

Mode 1: Informational Capture

Definition: The condition in which individuals and communities lose reliable access to accurate, comprehensible information about AI systems that affect their lives — either because that information is deliberately withheld, technically inaccessible, or systematically degraded by AI-generated noise in the information environment.

Mechanism: Informational capture operates through three sub-channels. Opacity by design: AI systems used in consequential domains — hiring, credit, healthcare, criminal justice — are frequently deployed without meaningful disclosure of their decision logic, training data characteristics, or performance limitations. Complexity as effective withholding: even when disclosure requirements exist, technical complexity functions as a practical barrier for affected individuals who lack the expertise to interpret what is disclosed. Epistemic saturation: AI-generated synthetic content floods the information environment in ways that make accurate signal increasingly difficult to distinguish from noise, degrading the epistemic commons on which political deliberation depends.^24,25

AI examples: Algorithmic hiring systems that produce rejection decisions without explaining which candidate attributes drove the outcome; AI credit scoring models whose internal weighting is protected as proprietary; large language model outputs that mimic authoritative prose while potentially containing errors that lay users cannot easily identify; synthetic media deployed in electoral contexts to misrepresent candidates.

Governance implications: Informational capture requires disclosure-oriented governance responses: mandatory explanation requirements for consequential AI decisions, interpretability standards that go beyond technical transparency to substantive accessibility, and independent monitoring of AI-generated information environments. The EM-IAF's Uncertainty Disclosure dimension addresses the system-level component; community information access (Section X) addresses the population-level component.

Mode 2: Economic Foreclosure

Definition: The condition in which AI-driven shifts in labor markets, skill requirements, or economic structures eliminate viable livelihood pathways for specific populations — without those populations having advance warning or meaningful access to transition support — such that their economic futures are effectively determined by forces they cannot influence or adapt to.

Mechanism: Economic foreclosure differs from ordinary unemployment in a specific way: it involves the elimination of categories of economic participation, not merely individual positions. When AI systems substitute for cognitive tasks that previously defined an occupation's skilled core — legal research, radiological interpretation, financial analysis — the economic foreclosure is occupational rather than individual. Workers face not a job search but a reorientation of economic identity, requiring new skills, credentials, and sometimes geographic relocation — each imposing costs that are unevenly distributed across income levels, ages, and family circumstances.^5,6

AI examples: Large language model systems that reduce demand for entry-level legal research, eliminating the apprenticeship pathway through which junior lawyers develop expertise; AI diagnostic systems that shift the skill premium in radiology without creating equivalent training pipelines; AI-enabled platform work that converts stable employment into precarious task-based arrangements without employer obligations for skill development or income continuity.

Governance implications: Economic foreclosure requires governance responses that address the transition gap: anticipatory skill investment rather than reactive retraining, portable benefit structures that survive employment relationship changes, and income support mechanisms designed for multi-year occupational transitions. Section VII-B establishes why income replacement alone is insufficient.

Mode 3: Political Exclusion

Definition: The condition in which individuals and communities lack effective channels to influence the AI governance decisions that shape their circumstances — whether because those decisions are made in technical or commercial domains without public participation mechanisms, or because participation mechanisms exist formally but are structured such that affected community input does not materially influence outcomes.

Mechanism: Political exclusion in AI governance operates differently from exclusion in traditional political arenas. The decisions that most immediately affect communities experiencing AI transitions — which AI systems are deployed in their workplaces, how systems determine hiring and credit decisions, what data is collected — are primarily made by private firms operating within commercial rather than democratic governance frameworks. These decisions are not subject to notice-and-comment rulemaking, electoral accountability, or other mechanisms through which affected populations can influence government decisions. Technical complexity further concentrates effective decision-making authority among those with specialized expertise, reducing the proportion of governance decisions accessible to democratically legitimate input.¹²

AI examples: Workplace AI monitoring systems deployed without worker input or negotiation; municipal predictive policing systems adopted through procurement processes that excluded community participation; AI-driven content moderation systems affecting civic discourse without accessible challenge mechanisms; AI systems in public benefit administration determining outcomes for vulnerable populations without accessible appeal processes.

Governance implications: Political exclusion requires structural governance responses: mandatory stakeholder participation requirements for high-stakes AI deployments, accessible challenge mechanisms that do not require legal representation, and regulatory frameworks that extend democratic accountability into commercial AI deployment decisions. The Assessment Charter's Stage 1 dispute process is an example of the accessibility standard the Foundation considers minimum.

Mode 4: Institutional Impotence

Definition: The condition in which nominally available accountability mechanisms — dispute resolution processes, regulatory bodies, legal redress pathways — lack the practical capacity to provide meaningful accountability for AI system harms, because those mechanisms are under-resourced relative to the scale of AI deployment, lack technical expertise to evaluate AI-specific claims, or face structural limitations that make accountability formally available but substantively inaccessible.

Mechanism: Institutional impotence differs from political exclusion in that it concerns the functional effectiveness of existing institutions rather than the formal availability of participation channels. Courts, regulatory agencies, and legal aid organizations may have nominal jurisdiction over AI-related harms while lacking the technical capacity to evaluate AI-specific evidence, the regulatory bandwidth to investigate at the scale of AI deployment, or the legal frameworks to assign liability in systems where harm is distributed across developers, deployers, and operators. The result is a governance gap in which accountability exists on paper while providing de facto impunity in practice.¹⁴

AI examples: Employment discrimination claims involving AI hiring systems that fail because plaintiffs cannot access the system's internal logic; privacy enforcement actions against AI data practices that are technically documented but legally ambiguous under existing statutory frameworks; local governments that adopt AI systems in criminal justice without staff capable of auditing or challenging outputs.

Governance implications: Institutional impotence requires capacity-building governance responses: investment in regulatory technical capacity, legal framework updates that address AI-specific accountability gaps, and institutional design that matches oversight capacity to the scale and pace of deployment. The EM-IAF's assessment methodology is designed in part to address the technical capacity gap for organizations without in-house AI evaluation expertise.

Mode 5: Temporal Lock-In

Definition: The condition in which current AI deployment decisions, infrastructure investments, or regulatory frameworks create path dependencies that foreclose future populations' capacity to make different choices — by establishing technical standards difficult to revise, concentrating commercial control in ways that make competition structurally impractical, or creating institutional arrangements that persist beyond the conditions that justified them.

Mechanism: Temporal lock-in is the least visible mode of agency loss because it operates prospectively rather than immediately — it constrains future populations rather than present ones. AI-specific mechanisms include: foundation model dependencies that concentrate future development around a small number of architectural choices; data concentration effects where organizations with large early datasets acquire self-reinforcing advantages; and governance frameworks calibrated to current AI capability levels without provisions for revision as capabilities change, creating regulatory structures that may be either inadequate or excessive under future conditions.^26,27

AI examples: Healthcare AI infrastructure that creates dependency on specific vendors' data formats, making future changes prohibitively expensive; AI regulatory frameworks enacted without sunset provisions or mandatory review mechanisms; international AI governance agreements reflecting current power distributions, making future rebalancing structurally more difficult; municipal AI procurement committing local governments to long-term vendor relationships without performance-based exit provisions.

Governance implications: Temporal lock-in requires reversibility-oriented governance responses: sunset provisions and mandatory review mechanisms in AI regulations, interoperability and portability requirements that preserve switching capacity, and procurement standards that build revision rights into long-term AI system contracts. Temporal agency (Dimension 5) tracks this mode specifically.

VI-B. Agency Enhancement — How AI Can Strengthen Rather Than Erode Agency

The five modes of agency loss in Section VI-A identify the pathways through which AI deployment threatens human agency. That analysis is necessary but asymmetric: a framework that exclusively models decline cannot serve as the basis for governance that preserves and strengthens agency, because it has no account of what strengthening looks like. This section identifies the mechanisms through which AI deployment can increase rather than decrease the five dimensions of agency, the conditions under which those mechanisms operate, and the governance implications of the asymmetry between erosion and enhancement dynamics.

The fundamental observation is that AI systems are not intrinsically agency-diminishing. They are capability-amplifying: they extend what individuals and institutions can do. Whether that amplification reduces or increases agency — whether it substitutes for human judgment and participation or augments the capacity for informed, effective participation — is determined by design choices, deployment contexts, and governance conditions, not by the technology itself. The following examples are not exhaustive; they are illustrative of the dimensions along which agency enhancement is documented or plausibly achievable.

Epistemic Enhancement: AI as an Access Equalizer

Access to accurate, understandable information has historically been constrained by economic barriers, technical expertise requirements, and geographic limitations. AI systems have genuine capacity to reduce all three constraints in ways that increase epistemic agency for populations that have historically had the least access to expert information.

Plain-language translation of legal, medical, and financial documents — converting technical language into forms accessible to non-specialist readers — is currently achievable with AI systems at quality levels that are useful if not perfect. For the approximately 54% of American adults who read below a sixth-grade level, AI-assisted plain-language translation of the AI system disclosures that the epistemic agency dimension requires could transform nominal transparency into substantive epistemic access. AI-powered question-answering interfaces on government and institutional websites could provide the interactive access to information that static disclosure documents do not — allowing affected individuals to ask the specific questions relevant to their circumstances rather than wading through general documentation that may not address them.

Multilingual access represents perhaps the clearest case of AI as epistemic equalizer. For linguistic minority populations — which disproportionately overlap with populations most exposed to AI-driven employment transitions — language barriers systematically limit access to governance information, legal documentation, and institutional resources. AI translation systems are not perfect, but they are substantially better than no translation, and they are continuously improving. Governance frameworks that require AI-assisted translation of consequential AI system disclosures into languages spoken by affected communities would substantially improve epistemic access for populations that formal transparency requirements alone do not reach.

Economic Enhancement: AI as a Transition Tool

The economic foreclosure mode of agency loss (Mode 2) emphasizes how AI eliminates occupational pathways. But AI systems can also create and expand pathways — particularly for populations whose economic options have historically been constrained by geographic isolation, physical limitations, or credential barriers that do not reflect genuine skill deficits.

AI-powered skills assessment and career guidance tools can identify transferable skills from prior employment that workers may not recognize as relevant to emerging occupations — reducing the information asymmetry that makes occupational transitions unnecessarily difficult. The dominant occupational transition challenge is not skill absence but skill recognition: workers who have developed project management, customer service, quality control, or process optimization skills in manufacturing contexts frequently possess skills that are genuinely valuable in service sector employment but cannot translate them into job market terms without guidance. AI systems trained on occupational competency data can perform this translation at scale, at low cost, and with the personalization that makes generic career counseling insufficiently useful for most workers.

AI-enabled remote work and gig economy platforms have expanded economic access for some populations — caregivers, people with disabilities, and rural workers for whom geographic proximity to employment was previously a binding constraint. The precarity concerns associated with gig platforms are real and documented (see Mode 2's treatment of platform work), but the expansion of access is also real. The governance challenge is not to eliminate the access gains while addressing the precarity problems but to preserve the former while solving the latter — an objective that requires both employment law reforms and the kind of community resilience infrastructure analyzed in Section X.

Political Enhancement: AI as a Participation Infrastructure

Democratic participation in AI governance is constrained by the same technical complexity that makes epistemic access difficult. The public comment processes through which affected communities are nominally consulted on AI governance decisions require participants to navigate technical documentation, understand regulatory terminology, and articulate concerns in the format that administrative processes recognize. These requirements systematically exclude the communities most affected by AI transitions — precisely the populations with the least institutional experience of formal regulatory participation.

AI-assisted participation tools can reduce these barriers in documented ways. Natural language interfaces that allow affected community members to describe concerns in ordinary language — and translate those concerns into the structured formats that administrative processes require — have been piloted in several regulatory contexts with positive results for participation diversity. AI-powered analysis of public comment submissions can identify thematic patterns across large volumes of responses — enabling regulatory agencies to understand community concerns at a scale that manual review cannot achieve without filtering out community voices that have been systematically underweighted. Whether these participation tools actually improve the influence of community input on governance decisions — as opposed to merely increasing the volume of participation — is an empirical question that the Political agency dimension (Dimension 3) is designed to track.

AI systems can also enhance the informational quality of democratic deliberation about AI governance itself — providing accessible explanations of technical AI concepts, comparative analysis of regulatory approaches in different jurisdictions, and plain-language summaries of proposed regulations that enable affected communities to participate in governance discussions without requiring technical expertise. This is the governance use case for the same AI capabilities that, in adversarial deployment contexts, produce the informational capture and epistemic saturation effects described in Mode 1.

Institutional Enhancement: AI as an Accessibility Tool

The institutional impotence mode (Mode 4) documents how accountability mechanisms that are formally available are often substantively inaccessible due to cost, complexity, and technical expertise requirements. AI systems can reduce each of these barriers for affected individuals seeking accountability for AI system harms.

AI-powered legal assistance — systems that can explain legal rights, identify relevant legal frameworks, draft complaint letters, and guide individuals through administrative challenge processes — has been documented to increase the rate at which individuals with AI-related grievances initiate formal challenge processes. The Millbrook County scorecard's finding that workers with AI-related grievances face a median resolution time of 14 months is partly a product of procedural barriers that discourage challenge initiation; AI assistance in navigating those barriers can increase challenge rates while reducing the legal expertise required. Equal access to AI legal assistance — ensuring that the populations most likely to be harmed by AI systems are not systematically disadvantaged relative to the organizations deploying those systems in their capacity to navigate legal processes — is a governance goal that the institutional agency dimension tracks.

AI-assisted monitoring of AI system compliance — using AI to audit AI — represents the most technically novel institutional enhancement pathway. If AI systems can identify when other AI systems are producing biased, erroneous, or undisclosed outputs at the scale and speed of AI deployment, they can provide the oversight infrastructure that human regulatory capacity cannot provide at equivalent scale. The EM Foundation's Assessment Charter and EM-IAF methodology are designed to provide the evaluation framework within which AI-assisted compliance monitoring would operate; the technical capacity for such monitoring is developing faster than the governance frameworks required to deploy it legitimately.

The Enhancement Condition: Design and Deployment Context

The cases above share a common structure: AI enhances agency when it is deployed with explicit design goals oriented toward epistemic access, economic transition support, democratic participation, or institutional accessibility — and when deployment is paired with governance mechanisms that ensure the enhancement functions as intended and reaches the populations it is designed to serve. The same underlying capabilities that produce epistemic enhancement (plain-language explanation) also produce informational capture (persuasive synthetic content). The same capabilities that produce economic enhancement (skills matching) also produce economic foreclosure (occupational substitution). The difference is design intent, deployment context, and governance accountability.

This conditional structure is the most important finding of the enhancement analysis: agency enhancement is not an automatic property of AI deployment but a governance achievement. It requires explicit goal-setting, design accountability, and measurement — precisely the governance infrastructure that the HAI framework is designed to provide. A governance framework that measures only erosion would fail to detect enhancement opportunities and would be unable to distinguish AI deployments that strengthen agency from those that erode it. The HAI's five dimensions are designed to be directionally neutral — capable of measuring increases as well as decreases in each dimension — so that governance attention can be allocated to both the erosion cases that require intervention and the enhancement cases that deserve replication.

VII. Social Cohesion, Institutional Trust, and Why the AI Transition Is Different

The social science literature on trust and social capital offers consistent findings: institutional trust, once eroded, rebuilds slowly; communities with higher generalized trust demonstrate consistently better governance outcomes, stronger economic performance, and greater resilience to collective problems.^21,22,23 The AI transition creates specific trust risks that existing social capital literature does not fully address, because the mechanisms by which AI systems erode trust are structurally different from the mechanisms that previous technology transitions employed.

The first mechanism is epistemic: AI systems can erode shared factual ground — the information commons on which political deliberation depends — at machine scale. The production of convincing synthetic content, combined with algorithmic personalization that reinforces existing beliefs rather than challenging them, creates conditions in which maintaining shared epistemic ground is structurally harder. This is not merely a disinformation concern; it is a democratic legitimacy concern, because political deliberation requires a shared factual substrate that AI systems can degrade at a speed print and broadcast media never approached.²⁴

The second mechanism is decisional opacity: AI systems make consequential decisions in ways that are not legible to the people affected by them. When a credit decision, hiring recommendation, or risk assessment is made by a system the affected person cannot interrogate, the experience of institutional illegitimacy is produced regardless of whether the decision is substantively defensible.²⁵ Opacity in consequential decisions is therefore not merely a fairness problem — it is a trust problem with cumulative effects on institutional legitimacy over time. The EM Foundation's Corroboration Standard addresses this at the answer level; the EM-IAF addresses it at the system level. Neither is sufficient without institutions capable of applying them consistently.

VII-A. Legitimacy and the Stability Threshold

Political legitimacy theory, from Weber's foundational typology through the contemporary social contract literature, converges on a counterintuitive finding: societies rarely become unstable because they become poor. Poverty is compatible with political stability, sometimes for long periods. Societies become unstable when a sufficiently large portion of the population concludes that the system no longer contains a viable future for them — that the institutions governing their lives are neither responsive to their interests nor capable of becoming so through legitimate means. The distinction is crucial. Poverty produces hardship. Illegitimacy produces revolutionary politics.^28,29

The historical record is consistent across cases that vary dramatically in their economic conditions. The French Revolution was not primarily a response to poverty — France in 1789 was not the poorest society in Europe, and the relevant comparison is not with absolute deprivation but with the perceived closure of legitimate pathways for the Third Estate. The precipitating conditions were epistemic (the fiscal crisis made the exclusion of commoners from governance information untenable), institutional (the Estates-General's convocation created a legitimacy contest the monarchy could not win), and temporal (reform proposals had been made and rejected for decades, foreclosing the gradualist alternative). What collapsed in 1789 was not the French economy. It was the population's belief that existing institutions could produce acceptable outcomes through legitimate processes.³⁰

Weimar Germany offers the sharpest lesson for AI governance purposes, precisely because the economic dimension was severe and the legitimacy dimension compound. The hyperinflation of 1923 and the depression of 1929 were not by themselves sufficient to produce the political catastrophe that followed. What made them catastrophic was their interaction with pre-existing legitimacy deficits: a republic that its founding demographic had never fully accepted as legitimate, combined with economic shocks that concentrated costs on precisely the constituencies most invested in the old order's legitimacy and most susceptible to the nationalist alternative's narrative of institutional betrayal. The economic crisis was real. But the mechanism of political collapse was legitimacy, not poverty.³¹

The Soviet collapse in 1991 is perhaps the most instructive case for technology transition governance because it occurred not in conditions of extreme poverty but in conditions of widespread perceived stagnation — the sense that the existing system had exhausted its capacity to deliver on its own promises. Gorbachev's glasnost accelerated collapse precisely because transparency revealed the gap between the system's legitimacy claims and its actual performance, and no institutional mechanism existed for the population to translate that information into legitimate institutional change. The Soviet state did not collapse because citizens became poor; it collapsed because enough citizens concluded simultaneously that the state had no legitimate path toward the future they were promised.³²

The Arab Spring of 2010–2012, though its outcomes varied dramatically across countries, shares the same underlying legitimacy dynamic. The self-immolation of Mohamed Bouazizi in Tunisia was not simply a response to economic grievance — it was a response to the specific experience of illegitimacy, the harassment of a licensed street vendor by officials whose authority he experienced as arbitrary and unaccountable. The subsequent spread of protest across the region documented not shared poverty but shared experience of governmental illegitimacy: the perception that existing institutions could not respond to legitimate concerns through legitimate processes, combined with the specific technologies that made shared illegitimacy visible and coordinable at speed.³³

The implications for AI governance are direct and underappreciated in the three primary narratives. AI can accelerate legitimacy loss through each of the mechanisms identified in the historical cases. It can erode the epistemic commons that legitimacy depends on, through synthetic content and personalized information environments that prevent the formation of shared factual ground. It can concentrate benefit-capture in ways that produce the specific experience of systemic exclusion — not poverty, but the perception that the system is producing outcomes for others while foreclosing them for you. It can generate decisional opacity at scale — hiring, credit, healthcare, criminal justice — that produces the Bouazizi experience of unaccountable authority over consequential decisions across millions of interactions simultaneously. And it can do all of this faster than the institutional adaptation mechanisms that historically absorbed legitimacy shocks have ever had to operate.

The connection between legitimacy and the five dimensions of human agency is therefore not metaphorical. Epistemic agency (Dimension 1) is the material prerequisite for the formation of legitimacy judgments — populations cannot conclude that the system is illegitimate based on accurate information if accurate information is systematically unavailable. Economic agency (Dimension 2) maps directly onto the perception of systemic inclusion or exclusion that drives legitimacy crises. Political agency (Dimension 3) is the institutional mechanism through which legitimacy concerns are supposed to translate into governance responses — its absence is the specific condition that converts grievance into revolutionary politics. Institutional agency (Dimension 4) is the operational test of legitimacy: whether accountability mechanisms function as claimed. And temporal agency (Dimension 5) addresses the forward-looking dimension of legitimacy — whether the system appears capable of producing acceptable futures, not only tolerable presents. Preserving human agency is, at the deepest level, the governance project of preserving legitimacy itself in the face of an AI transition that threatens each of its five material preconditions simultaneously.^34,35

VII-C. Relationship to Existing Governance Metrics

The Human Agency framework does not compete with existing governance metrics. It is complementary to them — designed to measure dimensions of social and political wellbeing that existing metrics either do not capture or capture only partially. Understanding precisely what established frameworks measure and where they reach their limits is necessary to situate the HAI's contribution accurately and to avoid the overclaiming that would undermine its credibility as a governance instrument.

Gross Domestic Product

GDP measures aggregate economic output — the market value of goods and services produced within a jurisdiction over a period. Its governance utility is real: sustained GDP growth is associated with expanded material living standards, increased public revenue for social investment, and the economic security that enables political stability. Its limitations for agency measurement are equally well-documented. GDP is indifferent to distribution: a society in which aggregate output grows while the distribution of that output becomes more concentrated registers GDP growth regardless of whether the typical person's economic agency has expanded or contracted. GDP does not measure the composition of economic activity — work that is demeaning, environmentally destructive, or produced under coercive conditions contributes to GDP equally with work that is meaningful, sustainable, and freely chosen. And GDP provides no information about the institutional conditions — accountability mechanisms, epistemic access, political voice — that the agency framework identifies as preconditions for the translation of economic output into genuine human welfare.

Employment and Labor Market Metrics

Employment rates and wage levels provide substantially more governance-relevant information than GDP alone: they measure whether economic output is being distributed through labor as well as capital, and whether the distributional structure of the economy is generating income security for working-age populations. Employment metrics are the most commonly cited indicators in AI transition governance debates, and the labor economics literature (Acemoglu and Restrepo, Frey and Osborne, Autor) uses employment as its primary dependent variable. The agency framework's contribution is to identify what employment metrics miss: as Section VII-B documents at length, employment is not only an income mechanism but an identity, purpose, community, and dignity framework. Employment metrics measure income; they do not measure whether employment provides the purposive structure that genuine economic agency requires. More specifically, they do not measure whether people have meaningful choice about the employment available to them — whether the options on offer are genuine choices or constrained selections from a foreclosed set.

Human Development Index

The UNDP's Human Development Index advances beyond GDP and employment by combining per capita income, educational attainment, and life expectancy into a composite measure of human development. The HDI's theoretical basis in Sen's capability approach makes it substantially more relevant to the agency framework than GDP alone: it recognizes that development is about expanding what people are able to do and be, not merely increasing aggregate output. The HDI's limitations for AI transition governance are two. First, it measures outcomes at a high level of aggregation that obscures the distributional patterns — by community, demographic group, and occupational category — that the agency framework treats as primary. Second, the HDI's dimensions are health, education, and income; it does not measure epistemic access to the AI systems that increasingly determine life outcomes, political voice in AI governance decisions, institutional recourse when AI systems cause harm, or temporal agency over the technological trajectories being set now for future generations. The HDI's framework and the HAI framework are complementary rather than competing: HDI measures development outcomes; the HAI measures the agency-preserving conditions under which development remains self-directed.²⁰

Social Capital Measures

Social capital measures — Putnam's civic participation indices, the World Values Survey's generalized trust measures, Fukuyama's institutional trust analysis — capture something important that economic metrics miss: the relational and associational infrastructure through which economic and political outcomes are produced. Communities with higher social capital demonstrate better governance outcomes, stronger economic resilience, and greater capacity for collective problem-solving. The AI transition's specific threat to social capital — through epistemic commons erosion, occupational community dissolution, and the substitution of algorithmic mediation for direct human relationship — is precisely the dynamic that Section VII documents. Where the HAI extends beyond social capital measures is in identifying the specific agency-preserving institutional conditions — epistemic access, political voice, institutional recourse, temporal reversibility — that social capital facilitates but does not itself guarantee. Social capital provides the relational infrastructure; the agency dimensions describe what that infrastructure must be used to accomplish if genuine agency is to be preserved.^22,23

Institutional Trust Metrics

Institutional trust measures — Edelman Trust Barometer, Gallup confidence-in-institutions surveys, OECD government trust indices — track the population's confidence in public and private institutions over time. These measures are directly relevant to the legitimacy analysis in Section VII-A: declining institutional trust is one of the most reliable leading indicators of political instability, and AI-specific sources of trust erosion (decisional opacity, epistemic commons degradation) are measurable through adapted versions of these instruments. The HAI's institutional agency dimension (Dimension 4) is related to but distinct from institutional trust: trust measures how the population evaluates institutions; Dimension 4 measures what institutional mechanisms are actually available and functional. A population can have high trust in institutions that are actually impotent — or low trust in institutions that are actually functional. Both the subjective and objective dimensions matter, and existing trust metrics capture only the former.

The Capability Approach (Sen and Nussbaum)

The capability approach, developed by Sen and elaborated by Nussbaum, provides the theoretical framework closest to the agency approach in this paper. Sen's argument — that development should be understood as the expansion of substantive freedoms to live the kinds of lives people have reason to value — is the direct conceptual ancestor of the HAI's attempt to define agency operationally across multiple dimensions. Nussbaum's specification of central human capabilities provides a list that substantially overlaps with the five agency dimensions: bodily health and integrity map onto economic agency; affiliation and political control map onto political and institutional agency; sense, imagination, and thought map onto epistemic agency.²⁰

The HAI framework's relationship to the capability approach is one of selective operationalization rather than theoretical competition. The capability approach provides the normative foundation; the HAI attempts to operationalize a subset of that foundation — the dimensions most directly threatened by AI transitions and most tractable for governance-relevant measurement. Two differences are worth noting. First, the HAI is explicitly designed for the AI transition context, which creates specific measurement priorities (epistemic access to AI systems, temporal reversibility of AI governance decisions) that the original capability framework does not address. Second, the HAI is designed to function as a governance metric for institutional use, requiring a level of operational specificity that the capability approach's philosophical generality does not provide. The HAI does not claim to replace the capability approach; it claims to operationalize a governance-useful subset of it for a specific historical and technological context.

The broader point about comparisons: the HAI is not superior to existing frameworks. It is differentiated — it measures things they do not measure, and it measures them at a level of specificity designed for the AI governance context. The appropriate governance response is not to choose between the HAI and existing metrics but to use them together: GDP and employment metrics for aggregate economic tracking; HDI for development outcomes; social capital and trust measures for relational and institutional health; and the HAI for the agency-preserving conditions that the other frameworks do not capture.

VII-B. Employment, Purpose, and Human Meaning

The governance literature on AI-driven employment change treats employment primarily as an income mechanism and a productivity determinant. This framing is analytically tractable and policy-relevant — it permits the labor economics literature to generate measurable predictions and policy prescriptions. But it is also radically incomplete. The sociological and psychological literature on work consistently documents that employment is not primarily an income mechanism for most people. It is an identity mechanism, a purpose structure, a community of belonging, and a dignity framework — each of these independently valuable and each independently at risk in AI-driven employment transitions that income replacement alone cannot address.

The sociological literature on work identity, developed from Durkheim's foundational analysis of occupational solidarity through contemporary studies of deindustrialized communities, documents that occupational identity is constitutive rather than instrumental for most workers. Miners, teachers, nurses, and machinists do not merely perform work for wages; they are miners, teachers, nurses, and machinists in a sense that organizes their self-understanding, their social relationships, and their sense of purpose in ways that are not reducible to the income stream the occupation provides.³⁶ The loss of manufacturing employment in communities like Youngstown, Ohio or Stoke-on-Trent, England produced social pathologies — increased substance abuse, family dissolution, suicide rates, civic disengagement — that preceded and persisted through economic recovery precisely because the income replacement that eventually occurred did not replace the identity, purpose, and community that the employment had provided.

Social identity theory, developed by Tajfel and Turner and subsequently applied extensively to organizational and occupational contexts, provides the psychological mechanism for this finding. Occupational group membership is among the most salient social identities for working-age adults — it shapes self-esteem, provides social comparison standards, and structures the in-group and out-group perceptions through which social belonging is organized.³⁷ The displacement of occupational identity through AI-driven unemployment is therefore not merely the loss of income; it is the disruption of a primary social identity, with psychological consequences that income replacement cannot address. The behavioral economics literature on the hedonic asymmetry of losses — losses loom larger than equivalent gains — suggests that the subjective experience of occupational identity loss is substantially more severe than the income calculation alone would predict.

The dignity dimension of employment is perhaps the most politically significant and most consistently underweighted in governance analysis. Empirical research on the relationship between employment status and dignity — drawing on both survey methodology and qualitative ethnographic work — finds that unemployment produces not merely financial stress but a specific experience of diminished social worth that persists independently of financial circumstances.³⁸ Paugam's comparative sociology of disqualification identifies paid work as the primary mechanism through which modern societies confer recognition of social contribution — and documents that the loss of this recognition, rather than the loss of income, is the primary driver of the political radicalization that accompanies long-term unemployment. Case and Deaton's "deaths of despair" literature confirms this finding in the specific American context of manufacturing employment decline: the mortality increases they document occur through mechanisms — substance use, suicide, violence — that are responses to experienced meaninglessness and lost social standing, not simply to financial deprivation.⁶

The purpose dimension of employment extends the dignity analysis. Psychological research on meaning-making consistently identifies purposive activity — work that is experienced as contributing to something beyond oneself — as a primary determinant of psychological wellbeing.³⁹ Employment, particularly skilled and recognized employment, provides this purposive structure for most working-age adults in ways that leisure, consumption, or passive income receipt do not. The governance implication is not that employment must be preserved at any cost — the historical transition literature documents that many displaced workers eventually find new purposive structures, occupational identities, and communities. It is that the transition period, during which these structures are disrupted before new ones are established, is itself a period of acute governance responsibility that income-focused frameworks systematically underweight.

The community dimension is the final element. Employment is not only individual; it is the primary organizational principle of many communities' civic and social life. Churches, labor unions, civic associations, and informal social networks are frequently organized around occupational communities — and when the occupational community dissolves, the civic infrastructure built around it dissolves with it. Putnam's analysis of social capital decline documents this dynamic in the context of manufacturing employment loss; the decline of union membership alone accounts for a substantial fraction of the measured decline in civic participation and associational life that his data documents.²² AI-driven employment displacement therefore poses risks not only to individual workers' income, identity, purpose, and dignity, but to the civic infrastructure of the communities affected — the institutional density that Section X identifies as a primary determinant of community resilience through technology transitions.

The governance implication that follows from this analysis is stark: preserving human agency through AI-driven employment transitions requires substantially more than preserving income. Income replacement — through unemployment insurance, universal basic income proposals, or comparable mechanisms — addresses one dimension of what employment provides while leaving the identity, purpose, community, and dignity dimensions unaddressed. A governance framework adequate to AI-driven employment transitions must grapple with the full sociology and psychology of work, not only its economics. This does not make income replacement unimportant — it makes it insufficient as a complete governance response. The Foundation's five-dimensional agency framework reflects this fuller analysis: Dimension 2 (economic agency) is defined as the capacity to pursue livelihood options and adapt skills, not merely to maintain income levels. The distinction is designed precisely to capture the gap between income replacement and genuine agency preservation in employment transitions.⁴⁰

VIII. Why the Transition Problem Is Institutional, Not Technical

The AI transition problem, as the Foundation understands it, is not primarily a technological problem. The technological questions — what AI systems can do, how capable they will become, how quickly — are inputs to the governance problem, not the governance problem itself. The transition problem is institutional: how do societies build and maintain governance institutions capable of assessing, accounting for, and constraining AI systems while those systems develop faster than institutions typically adapt?

This framing has three implications. First, governance must be built during the transition, not in response to harms that occur after capability thresholds are crossed. The interests surrounding powerful AI systems will systematically resist governance that constrains them — a structural dynamic documented across every previous technology transition and confirmed in the AI context by the behavior of major AI developers toward regulatory efforts.¹³ Second, governance institutions must be structurally resistant to capture. The Foundation's Assessment Charter devotes its entire Article IV to this requirement; the Anti-Capture Principle requires that the Foundation's assessment systems "resist capture from any direction — including industry, donors, governments, political movements, reviewers, providers, and the Foundation's own institutional preferences." Third, governance must be accountable to the people most affected by AI systems, who are disproportionately not the people currently building them.

The Transitional AGI Governance paper establishes the utility-first sequencing argument — prioritizing AI deployment in contexts where the primary measurable benefit is reduction of specific human vulnerability. This paper's contribution is complementary: the institutional conditions under which utility-first deployment actually reaches the communities it claims to benefit. Utility-first deployment without community-level epistemic access (Dimension 1) produces AI systems deployed in communities without the information required to evaluate whether the stated utility benefits are being realized. Utility-first deployment without institutional recourse (Dimension 4) produces AI systems whose failures harm the communities they purport to benefit without accountability.

IX. The EM Foundation's Position Within the Three-Narrative Landscape

X. Community Resilience — Why the Unit of Governance Analysis Matters

Macroeconomic analysis of technology transitions consistently overpredicts aggregate adaptation and underpredicts local disruption. The commuting-zone methodology used in Acemoglu and Restrepo's robot impact analysis is one of the most important methodological contributions to the transition literature precisely because it reveals what national-level analysis conceals: that the employment effects of automation are locally concentrated in ways that aggregate data obscures.⁵ A national employment gain that masks local employment loss in the community where a specific worker lives provides that worker with no agency over their actual circumstances.

The Foundation's Community Resilience Framework addresses this by identifying four community-level factors that determine whether a community can maintain agency through an AI transition, independent of national or sectoral aggregate outcomes.

Institutional density refers to the presence of local institutions — labor organizations, civic associations, libraries, community colleges, legal aid clinics — capable of providing information, retraining support, and political voice to community members affected by AI transitions. Communities with high institutional density have organizations that can identify when AI systems are being used in ways that harm community interests, provide information to affected members, and represent community interests in governance processes. Communities with low institutional density lack these capabilities regardless of what national-level governance frameworks exist on paper.

Information access refers to whether community members have access to accurate, understandable information about the AI systems affecting their employment, credit, healthcare, and civic participation. The EM-IAF's Uncertainty Disclosure and Transparency dimensions address this at the system level. Community information access requires that system-level transparency be complemented by community-level institutions capable of translating technical information into forms that affected community members can use to make meaningful choices.

Economic diversification refers to whether local economies have sufficient occupational and industrial diversity to absorb displacement from AI-affected sectors without requiring geographic relocation — which imposes costs (social network disruption, family separation, loss of community ties) that are real but not captured in standard labor market statistics.

Governance participation refers to whether community members have meaningful channels to influence AI governance decisions affecting their circumstances. This dimension is the political agency dimension (Dimension 3 of the five-dimensional framework) at the community level. Governance frameworks that make formal space for public comment while lacking mechanisms for community-level input to actually influence decisions satisfy the procedural form of governance participation without its substantive content.

X-A. Interaction Effects in the Resilience Model

The five community resilience variables are not independent — their interactions determine resilience outcomes in ways that make simple additive models insufficient. Three interaction patterns have particular governance implications.

Threshold effects: Resilience appears to require minimum threshold levels across all five variables rather than high levels in some variables compensating for very low levels in others. A community with extremely high institutional density — well-resourced civic associations, active labor organizations, accessible legal aid — but extremely low information quality will have organizations without the information they need to serve community interests effectively. A community with high economic diversity but negligible civic participation will have economic resources to absorb transition shocks but lack the political organization to ensure those resources are directed toward affected workers. Governance frameworks that concentrate support in the highest-scoring variables while neglecting threshold deficits in lower-scoring ones may achieve less resilience improvement than frameworks that target the lowest-scoring variables to bring them above minimum thresholds.

Sequencing effects: The order in which resilience variables are developed matters. Communities building resilience from very low baseline levels appear to benefit from prioritizing institutional density first — the organizational infrastructure that activates the other variables. Information quality improvements without institutions capable of distributing and acting on information do not translate into governance capacity. Civic participation improvements without local leadership to channel participation into effective advocacy may generate energy without direction. The sequencing implications for governance intervention: institutional density investment should precede or accompany information quality and participation initiatives rather than follow them.

Substitution limits: Economic diversity can buffer against acute transition shocks in ways that partially substitute for other resilience variables — a highly diversified local economy can absorb displaced workers without triggering the occupational identity crisis that monoeconomy communities face. But substitution has limits. Economic diversity does not substitute for institutional density in protecting political agency. It does not substitute for information quality in preserving epistemic access. And the buffer it provides against acute shocks does not prevent the slower erosion of governance participation that occurs when economic transitions are managed without community input. The governance implication: economic diversification policy is necessary but not sufficient for AI transition resilience, and must be accompanied by investments in the other four variables.

XI. Human Agency Index — A Research Program, Not a Ready Instrument

The purpose of a Human Agency Index (HAI) is to provide a longitudinal, community-level measure of whether the AI transition is preserving or eroding the five dimensions of human agency identified in Section VI. Existing surveys and indices — the OECD Better Life Index, the Gallup World Poll, the Social Progress Index — measure related constructs but do not specifically track the agency dimensions most relevant to the AI transition context. The HAI would be designed to fill this gap.

The HAI would require baseline measurement before significant AI deployment in a given community or sector — this is the temporal prerequisite that makes before-and-after comparison possible and that is unavailable retrospectively. This requirement mirrors the IAF Validation Roadmap's gate conditions for the Standard Benchmark: the instrument must be designed and deployed before the phenomenon it measures is fully underway, or the baseline from which change is measured does not exist. The HAI's most important measurement challenge is therefore not technical but temporal — establishing baselines now, before the AI transition in specific communities and sectors advances to the point where pre-transition measurement is no longer possible.

Four measurement domains: subjective epistemic access (do community members feel they understand AI systems affecting them?); objective institutional recourse (what accountability mechanisms are formally available and substantively accessible?); economic adaptive capacity (what retraining, income support, and labor market adjustment resources are available and utilized?); and governance participation (what formal and informal channels exist for community input to AI governance decisions, and is that input documented as influencing outcomes?). Each domain requires both objective and subjective measures, because the experience of agency is not fully captured by the formal availability of agency-preserving institutions.

XI-A. Example Scorecard — Illustrative Community Application

The following illustrative scorecard demonstrates how the HAI framework would be applied to a hypothetical mid-sized regional community ("Millbrook County") with a significant manufacturing employment base experiencing AI-driven workplace automation. Scores on a 0–10 scale are illustrative, not empirically derived. They are intended to demonstrate that the five dimensions can move independently — and to make visible the specific pattern in which a community can experience rising productivity while losing agency in dimensions that income statistics alone would not reveal.

The scorecard's most important finding is not the composite score but the pattern of dimension scores. Millbrook County has stable aggregate employment and measured productivity growth — conditions that would register as a "successful" AI transition in most macroeconomic frameworks. The HAI reveals that this aggregate picture masks severe erosion across three of the five agency dimensions. Communities and workers in Millbrook County cannot meaningfully understand, influence, or seek redress for AI systems that are materially shaping their employment circumstances. The aggregate economic indicators are improving while the conditions for democratic accountability over the AI transition are deteriorating.

This is the central measurement contribution that the HAI is designed to make: making visible the divergence between aggregate outcome metrics and agency-preserving conditions. A society can experience rising productivity while losing agency. Standard economic measurement does not detect this divergence. The HAI is designed to detect it specifically, providing the governance signal that macroeconomic measurement systematically suppresses.

XI-B. Scoring Methodology

Each dimension would be scored through a combination of objective institutional measures (formal availability of mechanisms) and subjective survey measures (experienced access and effectiveness). The composite score would weight the five dimensions initially equally, with the explicit acknowledgment that future research may provide empirical grounds for differential weighting — for example, if evidence establishes that epistemic access erosion systematically undermines all other dimensions, a weighted scheme that reflects this priority may be theoretically superior to equal weighting. The HAI research program would treat weighting methodology as an open empirical question, not a foundational assumption.

XI-C. Measurement Challenges and Validity Concerns

Construct Validity

Construct validity asks whether a measurement instrument actually measures the construct it claims to measure. For the HAI, construct validity requires demonstrating that the five dimensions — epistemic, economic, political, institutional, and temporal — are the right dimensions; that the indicators chosen for each dimension genuinely capture the dimension rather than a related but distinct concept; and that the composite score aggregates dimension scores in a way that reflects the underlying construct rather than an artifact of measurement choices.

Each of these requirements faces specific challenges. The choice of five dimensions reflects theoretical judgments about which aspects of agency are most governance-relevant — judgments that can be contested from within the capability approach (which might prioritize different dimensions), political philosophy (which might weigh procedural dimensions differently from substantive ones), and the empirical literature on wellbeing (which might include dimensions the HAI does not). The indicator choices face the classic social science problem of operationalization: retraining program completion rate, used as an indicator of economic adaptive capacity in the Millbrook County scorecard, measures something related to economic agency without measuring it directly. And equal weighting of the five dimensions is an explicit assumption rather than an empirically derived one — a reasonable starting point but not a theoretically motivated endpoint. Future validation work must test alternative operationalizations against the original and demonstrate convergent and discriminant validity across measurement approaches.

Cross-Cultural Variation

The five dimensions of agency reflect conceptual frameworks — particularly the capability approach and liberal democratic political theory — that are products of specific intellectual and cultural traditions. The degree to which these dimensions translate across the full range of political, economic, and cultural contexts in which AI transitions are occurring is an empirical question that the current framework does not answer.

Specific concerns: the political agency dimension assumes the existence of meaningful democratic participation channels as a relevant reference point — but in political systems where such channels are structurally absent or formally present but functionally empty, the dimension's measurement may produce results that are systematically different in meaning across political contexts. The epistemic dimension's focus on individual information access reflects an individualist epistemological framework that may not translate into contexts where knowledge is primarily communal and authoritative sources are institutional rather than individual. The temporal dimension's emphasis on reversibility may conflict with governance traditions that prioritize stability and continuity over adaptability. These concerns do not invalidate the framework — they identify the scope conditions that must be specified and tested before cross-cultural comparisons are valid.

Weighting Disputes

The composite HAI score requires a weighting scheme for the five dimensions. Equal weighting is the most defensible starting assumption in the absence of strong theoretical or empirical grounds for differential weighting — but it is almost certainly not the empirically correct scheme. There are plausible theoretical grounds for epistemic priority (Dimension 1 as the foundation for the others) that would argue for higher epistemic weighting. There are plausible empirical grounds, drawing on the legitimacy literature, for political priority (Dimension 3 as the mechanism through which other dimensions are protected). There are plausible practical grounds, drawing on the employment literature in Section VII-B, for economic priority (Dimension 2 as the most immediately experienced dimension for most populations). Governance uses of the HAI score — particularly any use that triggers policy responses or resource allocation decisions — require explicit justification of the weighting scheme and demonstrated robustness to alternative weighting assumptions. The Foundation does not claim equal weighting is correct; it claims it is the appropriate default while research on weighting is conducted.

Survey Bias

The HAI's subjective measurement domains — do community members feel they understand AI systems affecting them? do they experience governance participation as meaningful? — are subject to the full range of survey measurement biases. Social desirability effects may cause respondents to report higher epistemic access than they actually have. Framing effects may cause the same underlying experience to register differently depending on how questions are worded. Reference group effects — where respondents compare their agency experience to what they perceive as normal in their community rather than to an objective standard — may cause identical levels of actual agency to register as high or low depending on local expectations. Longitudinal surveys face attrition bias, with the populations most experiencing agency loss being potentially less likely to participate in follow-up measurement. The HAI research program must incorporate established survey methodology practices — cognitive interviewing, split-ballot testing, attention to non-response patterns — to manage these biases without eliminating the subjective measurement domains that capture dimensions of agency not accessible through objective indicators alone.

Political Misuse

Any governance metric that produces scored rankings is susceptible to political misuse: governments seeking to demonstrate favorable outcomes may selectively report, game the indicators, or manipulate the measurement conditions. The HAI is specifically susceptible to this risk because its policy salience — if adopted as a governance standard — creates incentives for score manipulation that do not exist for an academic research instrument. The structural protections required include: independent administration of measurement rather than government self-reporting; transparent methodology that allows independent replication; indicator design that is resistant to gaming (preferring objective institutional measures that require substantive reform over perceptible measures that can be manipulated through reporting); and explicit acknowledgment that indicator performance and underlying agency conditions can diverge. The HAI research program treats these structural protections as design requirements from the outset, not as optional refinements.

XI-D. Agency as a Potential Leading Indicator

Most governance measurement systems are designed to detect harm after it has occurred. Employment statistics identify displacement after workers have lost jobs. Institutional trust surveys identify erosion after trust has declined. Political participation measures identify disengagement after citizens have withdrawn. Each of these is a lagging indicator — useful for documenting what has happened but not for anticipating what is coming. The possibility that human agency measures could function as leading indicators — detecting governance stress before it manifests as visible harm — is a hypothesis worth examining carefully, though one that must be stated with appropriate caution.

The theoretical basis for the leading indicator hypothesis derives from the legitimacy literature reviewed in Section VII-A. The historical cases — France 1789, Weimar Germany, Soviet collapse, Arab Spring — all exhibit a consistent pattern: legitimacy erosion preceded visible political instability by periods measured in years, and that legitimacy erosion was detectable in principle from declining civic participation, eroding institutional trust, and withdrawal from formal political engagement before the structural crisis that these trends eventually produced. If agency dimensions are the material preconditions for legitimacy — as Section VII-A argues — then systematic decline in agency measures may anticipate legitimacy crises before those crises manifest as the political instability that lagging indicators would detect.

Specific agency dimensions may have differential leading indicator value. Epistemic access (Dimension 1) may be the earliest detectable signal: information environment degradation is measurable — through survey-based epistemic confidence measures, tracking of information source diversity, monitoring of AI-generated content prevalence — before it produces the downstream effects on political deliberation and institutional trust that become visible in governance outcomes. Political agency (Dimension 3) may provide intermediate-term signals: declining governance participation, withdrawal from public comment processes, and decreasing rates of challenge to AI-related institutional decisions are detectable before the political radicalization that participation decline eventually enables. Institutional impotence (Mode 4 in Section VI-A) may provide near-term signals: increasing rates of abandoned institutional challenge, growing average resolution times for AI-related disputes, and declining rates of successful redress are measurable indicators of institutional failure that precede the broader trust collapse that institutional impotence eventually produces.

Three important caveats are required. First, the leading indicator hypothesis is theoretical rather than empirically demonstrated. No longitudinal study has tested whether declining HAI scores predict political instability, and the data required for such a test does not exist because baseline HAI measurement has not yet been conducted. The hypothesis is plausible and theoretically grounded; it is not established. Second, even if agency dimensions are leading indicators, the lead time may be too short for governance response — the relationship between epistemic commons degradation and political crisis may involve dynamics that leave too little time for institutional adaptation even with early warning. Third, the leading indicator framing risks governance misuse: if declining agency scores trigger automatic policy responses, the political incentives for gaming the measures increase proportionally. The Foundation endorses further research into the leading indicator hypothesis while recommending that any governance use of HAI scores be paired with independent verification and resistance to gaming.

What can be said with confidence is that agency measures are more temporally proximate to the conditions that produce governance outcomes than the outcome measures themselves. If the theoretical framework in this paper is correct — that agency preservation is the governance mechanism through which legitimate technological transitions are produced — then agency measures should predict governance outcomes better than lagging indicators do. Testing this prediction is a primary research agenda item for the HAI program, and the Foundation treats it as a foundational empirical question rather than an established result.

XII. AI Impact Observatory — Data Infrastructure for a Transition Without Adequate Measurement

The AI Impact Observatory (AIO) is proposed as an independent, publicly accessible data infrastructure for tracking the real-world effects of AI deployment across communities, sectors, and demographic groups. The fundamental measurement problem the AIO addresses is this: AI systems are being deployed in high-stakes contexts — hiring, credit, healthcare, criminal justice, education — and the causal effects of those deployments on the people affected are not systematically tracked by any public institution. Researchers study individual cases; investigative journalists document specific abuses; academic audits reveal disparities in specific systems. But the aggregate, longitudinal, cross-sector data infrastructure required to track AI's distributional effects does not exist.

The governance challenge for such an observatory is significant and should not be understated. The data required to track AI impacts is largely held by the commercial entities deploying AI systems, who have limited legal obligation to share it and substantial commercial incentives not to. Building the observatory would require either regulatory data access requirements — which face political resistance from AI industry stakeholders — or voluntary data-sharing agreements that are fragile in the absence of legal obligation. The Foundation identifies this governance challenge not to dismiss the proposal but to situate it accurately: the AIO is a medium-term institutional goal that requires regulatory development, not only Foundation initiative.

The AIO's minimum viable form, achievable within the Foundation's current institutional capacity, would aggregate publicly available data: labor market statistics disaggregated by AI-intensive sectors and geographic units; documented AI system audits and bias findings; litigation records involving AI system harms; and survey data on public trust in AI-using institutions. This minimum viable form provides substantially more analytical value than the current fragmented evidence landscape, while the longer-term data access requirements are being established through regulatory and advocacy channels.

XIII. Governance Principles That Follow from the Human Agency Framework

This paper does not propose a comprehensive AI governance regime — that is beyond its scope and would require engagement with specific legal, regulatory, and political contexts that vary across jurisdictions. Five governance principles follow from the human agency framework and are consistent with the Foundation's existing documents. These are not new principles; they are applications of the human agency framework to governance decisions that are currently being made.

Governance must be built before harm, not in response to it. The historical record and the structural logic of technology transitions both support this principle. Post-hoc governance must contend with established commercial interests and path dependencies that pre-hoc governance can shape.¹³ The current moment — when AI capabilities are substantial but not yet so dominant that the interests surrounding them can effectively prevent governance — is the window for building institutions that will matter when capability levels are higher.

Distributional analysis must accompany aggregate analysis. Every major AI governance decision should be accompanied by analysis of its distributional effects across demographic groups, occupational categories, and geographic communities. Governance frameworks that focus exclusively on aggregate outcomes treat the unequal distribution of transition costs as acceptable if the aggregate numbers look favorable — which is the characteristic failure mode of every previous inadequate technology transition governance.

Epistemic access is a precondition for other forms of agency. The first dimension of human agency is prior to the other four. Individuals cannot exercise economic, political, institutional, or temporal agency in relation to AI systems they cannot understand or obtain accurate information about. Transparency requirements for consequential AI systems — those affecting employment, credit, healthcare, housing, education, and civic participation — are therefore a prerequisite for human agency, not merely a desirable feature. The EM-IAF's Uncertainty Disclosure dimension provides a system-level measurement instrument for this requirement.

Accountability mechanisms must be accessible to the people most affected. Governance institutions accessible only to organizations with legal resources, technical expertise, or political connections do not provide accountability for the individuals and communities most affected by AI deployments. The Assessment Charter's dispute resolution process is designed with accessibility as a constraint, not an afterthought — Stage 1 review requires no legal representation and carries no filing fee. The same accessibility standard should apply to every AI governance accountability mechanism.

The temporal dimension of agency requires reversibility. Governance decisions made now that foreclose future generations' choices about AI systems violate the fifth dimension of human agency regardless of their near-term benefits. This argues for governance approaches that are revisable as evidence accumulates — that avoid locking in technical architectures, commercial structures, or regulatory frameworks that are difficult to alter as understanding of AI impacts improves.

XIV. The EM Foundation Governance Stack

The governance framework described across this paper and the Foundation's companion documents is not a collection of independent instruments. Each element is functionally dependent on the others in a sequenced architecture that can be described as a governance stack — a layered structure in which each layer provides the preconditions that make the next layer possible. Understanding this architecture is essential both for evaluating the completeness of the Foundation's governance approach and for identifying where institutional gaps create risks that no individual instrument can address.

The stack's structure illustrates a governance failure mode that is not immediately obvious from examining individual instruments in isolation: a gap at any layer propagates downward. Assessment without corroboration produces evaluations that cannot be independently verified. Corroboration without a trust ledger produces dispersed findings that cannot accumulate into longitudinal records. A trust ledger without accountability produces records that have no governance consequence. And accountability mechanisms that do not connect to the human agency framework produce formal redress that leaves the legitimacy preconditions for social stability unaddressed. The governance stack is therefore an argument for comprehensiveness — not merely for any individual instrument's quality.

Critiques and Alternative Frameworks

The human agency framework advanced in this paper faces a set of serious objections from scholars and practitioners operating within alternative frameworks. This section presents each objection in its strongest form and responds carefully. The Foundation does not dismiss these objections; they identify genuine tensions in the agency framework that future research must address.

How the Framework Could Be Used

The human agency framework is offered as a research program and governance orientation, not as a ready implementation manual. The specific instruments — Human Agency Index, AI Impact Observatory, community resilience assessments — require methodological development before they can function as operational governance tools. This section describes how the framework's conceptual structure and existing proxy measures could be used now, before the full instrument suite is developed, by different categories of actors with different governance responsibilities and capacities. The goal is to demonstrate practical relevance without overclaiming operational readiness.

Local Governments

Local governments are frequently the first institutions to encounter AI transition effects and the least resourced to respond to them. Displaced workers interact with local government through unemployment services, workforce development programs, and community college systems. Local governments are also frequently excluded from the AI procurement and deployment decisions that drive the transitions they must then manage.

Practical uses of the framework at the local level: First, the community resilience model (Section X) can be used as a diagnostic tool for assessing a community's baseline position before major AI deployments in local industries — identifying which of the five resilience variables are below threshold and which governance investments would most improve resilience capacity. Second, the HAI scorecard structure (Section XI-A) provides a template for the kind of monitoring local governments could conduct using existing data sources — labor market statistics, community college enrollment, public comment participation rates — without waiting for the full HAI instrument to be developed. Third, the five modes of agency loss (Section VI-A) provide a typology that local governments can use to audit existing AI systems in their own operations — checking whether city government use of AI in permit processing, benefit determination, or code enforcement produces any of the five agency failure modes for residents.

National Governments

National governments have the regulatory authority and data access that local governments lack, but may be further from the community-level effects that the framework identifies as primary governance units. The framework's practical contributions at the national level are primarily in two areas.

First, distributional analysis requirements: the framework's argument that aggregate outcome analysis systematically obscures the community-level effects of AI transitions provides justification for regulatory requirements that accompany aggregate economic reporting with disaggregated community-level analysis. National governments adopting AI impact reporting frameworks could use the five agency dimensions as reporting categories, requiring AI developers and deployers to report not only aggregate benefit estimates but epistemic access provisions, political participation mechanisms, institutional recourse pathways, and reversibility safeguards. Second, the leading indicator function (Section XI-D) suggests a governance use for the HAI that national governments may be best positioned to operationalize: longitudinal monitoring of agency dimensions in communities experiencing significant AI-driven employment transitions, to provide early warning of governance stress before it produces the legitimacy crises documented in Section VII-A.

Non-Governmental Organizations

NGOs operating in the AI governance space — civil society organizations, labor advocacy groups, digital rights organizations, community organizing nonprofits — are often best positioned to conduct the independent monitoring and advocacy functions that the framework identifies as necessary. Their practical uses of the framework are primarily adversarial and accountability-oriented: using the five modes of agency loss as a documentation framework for specific AI system impacts; using the community resilience model to identify communities at particular risk of transition failure; and using the HAI structure as an advocacy tool to argue for the measurement infrastructure that the framework requires.

The framework also provides NGOs with a conceptual vocabulary for bridging between the technical AI governance discourse — which tends to focus on bias, safety, and capability — and the community organizing discourse — which tends to focus on jobs, dignity, and political voice. The five-dimensional agency framework and the five modes of agency loss provide a structured account of how these concerns connect, which may be useful for coalition-building across organizations that currently operate in separate discursive registers.

Academic Researchers

For academic researchers, the framework provides a research agenda rather than an implementation pathway. The open questions identified in Section XVII and the measurement challenges in Section XI-C define the empirical work that the framework requires to become an operational governance instrument. Three research priorities are most immediate.

First, construct validity testing: developing and testing alternative operationalizations of the five agency dimensions against the original HAI specification, using established psychometric methods to assess convergent validity (do alternative measures of the same dimension correlate strongly?) and discriminant validity (do measures of different dimensions correlate weakly?). Second, cross-cultural calibration: testing whether the five dimensions are interpreted consistently across cultural contexts, and whether the indicators that capture each dimension in one context function equivalently in different political, economic, and cultural settings. Third, leading indicator validation: assembling historical data on community-level agency indicators (where available from existing surveys and institutional records) and testing whether patterns of agency decline precede and predict the governance outcomes — declining civic participation, eroding institutional trust, political radicalization — that the theoretical framework predicts.

Corporations

Corporations developing and deploying AI systems have both the greatest capacity to affect the five dimensions of human agency and the greatest commercial incentive to resist governance frameworks that constrain that capacity. The framework's practical value for corporations is primarily in two non-adversarial use cases.

First, internal risk assessment: the five modes of agency loss provide a structured framework for evaluating AI systems before deployment for the governance risks they create. Corporations with sophisticated legal and compliance functions can use the typology to identify whether a specific AI deployment creates informational capture risks (inadequate disclosure provisions), economic foreclosure risks (transition support inadequacy), political exclusion risks (no worker input mechanism), institutional impotence risks (grievance process inadequacy), or temporal lock-in risks (insufficient reversibility provisions). This pre-deployment assessment does not require accepting the full HAI framework; it requires only taking seriously the possibility that AI-related governance failures create legal, reputational, and regulatory risks that internal assessment should identify. Second, stakeholder reporting: the five agency dimensions provide a reporting structure that goes beyond the current ESG framework's treatment of AI — which is typically limited to bias and safety — to address the governance and democratic accountability concerns that the framework identifies as primary.

Community Organizations

Community organizations — neighborhood associations, religious institutions, mutual aid networks, local advocacy groups — are not typically actors in AI governance discourse. The framework argues that they should be, because they are the institutions best positioned to detect the community-level effects that the framework identifies as primary governance concerns. Their practical use of the framework is primarily in three areas: documenting agency loss in specific community contexts using the five-mode typology; representing community interests in formal governance processes using the language and structure of the five-dimensional framework; and building the institutional density that the community resilience model identifies as a primary determinant of transition resilience. The Foundation's commitment to accessible governance processes — Stage 1 review without legal representation, accessible dispute resolution — reflects the judgment that community organizations are a necessary actor in AI governance, not a supplementary one.

Strongest Criticisms of the Framework

The previous section addressed objections to the priority of agency relative to welfare, growth, existential risk, and alternative political frameworks. This section addresses criticisms that target the framework itself — its conceptual specificity, its measurability, its economic assumptions, its cultural transferability, and its relationship to existing measurement approaches. These criticisms are presented in their strongest form; responses are offered where they are available and shortcomings are acknowledged where they are not.

What This Paper Does Not Claim

This paper does not claim that the replacement narrative is more accurate than the utopia narrative, or vice versa. The empirical question of net AI employment effects over the next 20–30 years is genuinely uncertain and will be determined by factors — the pace of AI capability development, the institutional responses of governments and civil society, the investment decisions of AI developers — that cannot be predicted with confidence from present evidence.

This paper does not claim that the Human Agency Index and AI Impact Observatory can be easily built. Both proposals face substantial methodological, political, and financial obstacles that the paper acknowledges and does not have solutions to. They are offered as research agenda items, not as ready instruments.

This paper does not claim that the five-dimensional definition of human agency is the only defensible operationalization. Scholars working in the capabilities literature (Sen, Nussbaum), the political philosophy literature (Pettit, Young, Frankfurt), and the labor economics literature would define the concept differently — and some would argue that the operationalization here is either too narrow or too broad for rigorous measurement. The definition is offered as a governance-useful operationalization, not as a contribution to the philosophical literature on agency.

This paper does not claim that historical technology transitions are directly analogous to the AI transition. AI may be advancing faster than any previous general-purpose technology; may displace cognitive tasks in ways previous automation did not; and may produce more extreme distributional outcomes due to the geographic and demographic concentration of AI development. These disanalogies are genuine limitations on the conclusions that can be drawn from historical precedents.

This paper does not claim that international dimensions are adequately addressed. The governance frameworks discussed are primarily national or regional. The distributional effects of AI transitions differ substantially across countries — AI may displace workers in lower-income countries through automation of previously offshored tasks while generating productivity gains primarily in higher-income countries. A comprehensive governance framework would need to address these international dimensions.

Open Questions

Known Limitations

The empirical evidence base is uneven. Evidence on labor displacement effects (Acemoglu and Restrepo, Frey and Osborne) is methodologically stronger than evidence on labor creation effects from AI, because displacement is more precisely measurable than the diffuse creation of new occupational categories. The paper's treatment of both evidence bases as roughly comparable in strength may overstate the evidentiary support for the utopia narrative's empirical foundations relative to the replacement narrative's.

The community resilience framework lacks empirical validation. The four community resilience factors are identified from the governance literature and historical cases, not from systematic empirical research on AI transition effects at the community level. They are plausible and grounded in analogous literatures, but they have not been tested against actual AI transition data because adequate data does not yet exist. The AI Impact Observatory proposal is designed specifically to address this gap.

The governance principles in Section XIII are stated at a high level of generality. The paper does not provide implementation criteria — the observable conditions that would allow an independent assessor to determine whether a specific governance framework satisfies each principle. The Transitional AGI Governance paper's treatment of its seven principles, which includes implementation criteria for each, represents the standard the Foundation aims for. Future iterations of this work should provide comparable specificity.

The legitimacy analysis in Section VII-A is primarily historical and qualitative. The four historical cases — France 1789, Weimar Germany, Soviet collapse, Arab Spring — are suggestive of a general pattern but do not constitute a systematic empirical test of the legitimacy threshold argument. The causal mechanisms proposed are theoretically grounded but empirically unverified in the specific AI context. The Foundation presents this analysis as hypothesis-generating rather than hypothesis-confirming.

The framework comparison in Section VII-C is selective rather than comprehensive. The governance metrics compared — GDP, employment, HDI, social capital, institutional trust, capability approach — are the most commonly cited in AI governance literature but do not exhaust the space of relevant measurement frameworks. Wellbeing economics (life satisfaction indices), democratic quality indices, and social cohesion measures each capture dimensions relevant to the governance questions this paper addresses. A comprehensive comparison would require substantially more space than this paper provides.

The sociology-of-work analysis in Section VII-B draws primarily on Western European and American research contexts. The relationship between employment, identity, purpose, and community may differ substantially across cultural and institutional contexts. The AI employment transition will affect communities in vastly different cultural settings, and governance frameworks adequate to those settings may require different sociological foundations than those this paper employs.

Conclusion

The three dominant narratives about artificial intelligence are each partially right, each partially wrong, and each incomplete in a way that has specific governance costs. The replacement narrative accurately identifies that AI transitions impose real costs on real communities but systematically aggregates those costs in ways that obscure who bears them. The utopia narrative accurately identifies that AI can generate substantial human benefits but consistently underweights the institutional conditions required to realize those benefits equitably. The doom narrative accurately identifies that AI development poses genuine risks but sometimes focuses on speculative long-run scenarios while current harms proceed without adequate accountability. And cutting across all three, the concentration variant identifies the structural condition that makes each narrative's worst outcome more likely and each narrative's best outcome less achievable.

The missing element in all three narratives is human agency — the structured capacity of individuals, communities, and institutions to make meaningful choices about their futures in the context of AI transitions. The five-dimensional operational definition proposed in this paper — epistemic, economic, political, institutional, and temporal — is designed to make agency a governance metric rather than a rhetorical aspiration. The Human Agency Index and AI Impact Observatory proposals are designed to provide the data infrastructure that a governance framework focused on agency preservation requires.

The Foundation's governance position is not between the three narratives. It is beneath them — providing the trust infrastructure, assessment methodology, accountability mechanisms, and anti-capture architecture that each narrative independently requires to achieve its own stated goals. The governance stack described in Figure 5 demonstrates how these instruments fit together: not as a collection of independent proposals, but as a functional dependency chain in which each layer provides the preconditions for the layer below. The scenario that all three narratives claim to want — a high-capability AI transition that broadly benefits humanity — is achievable. Whether it is achieved depends primarily on governance decisions being made now, before AI capability levels and the interests surrounding them make the governed transition structurally harder to build toward. Legitimacy is the hidden variable in every technology transition. Preserving it, through the preservation of human agency in each of its five dimensions, is the governance challenge of our moment.