Alignment Through
Mutual Dependence

Paper structure

Alignment must become a living public institution

The technical alignment problem is real: advanced AI systems can fail, misgeneralize, deceive evaluators, amplify harmful capabilities or act too quickly for human intervention. The 2026 International AI Safety Report emphasizes that agentic systems raise the consequences of error because they can perform multi-step actions autonomously, while current safeguards remain imperfect.¹ Yet technical control is only one part of the problem.

An AI system can behave exactly as its operator intends and still be socially misaligned. It may faithfully maximize advertising revenue, labor substitution, surveillance capacity, military advantage or asset returns. From the operator's perspective the system is aligned. From the perspective of workers, communities, future generations or democratic self-government, it may be profoundly misaligned.

The question is therefore not only whether an AI pursues a goal. It is whether the process that generated the goal is legitimate, plural, revisable and connected to those who bear the consequences. NIST's AI Risk Management Framework already treats governance, mapping, measurement and management as continuous lifecycle functions and calls for engagement with affected actors and feedback about unanticipated impacts.² Commonsent extends that logic from organizational risk management to a societal operating architecture.

Five propositions

Commonsent reframes alignment from "How do we make the machine obey?" to "How do humans and machines remain mutually corrigible within institutions that no narrow group can permanently capture?"

Why current approaches to alignment are incomplete

2.1 Model alignment is not system alignment

Training techniques such as reinforcement learning from human feedback, constitutional rules, red teaming and monitoring can reduce harmful behavior. They do not determine who owns the system, who chooses its permitted purposes, who receives its productivity gains or whether affected people can appeal its decisions. A safe model inside an extractive institution can make that institution more capable.

This distinction becomes more important as AI moves from answering questions to taking actions. An agent can schedule, purchase, negotiate, write software, manage infrastructure or coordinate other agents. Each action inherits permissions from an institutional environment. Runtime governance - identity, authorization, spending limits, tool access, audit logs, human escalation and revocation - becomes as important as model behavior.

2.2 Alignment to a static value set is brittle

Human societies disagree. They also learn. New technologies create situations that existing moral vocabularies did not anticipate. A fixed constitution can provide a rights floor, but it cannot settle every future tradeoff. Alignment therefore needs two layers: relatively stable protections and continuously revisable policies.

Collective Constitutional AI demonstrated that public input can be gathered and incorporated into model behavior, using a process involving roughly one thousand Americans.³ This is important evidence of feasibility. It is not a complete democratic institution. A one-time sample cannot represent all populations, changing conditions or context-specific disputes. Commonsent turns such exercises into standing, versioned and auditable processes.

2.3 Aggregation is not neutral

There is no universally correct method for converting millions of diverse preferences into one coherent instruction. Social-choice research shows that aggregation rules necessarily embody tradeoffs and can produce incompatible results under different assumptions.⁴ Recent work specifically argues that social choice should guide AI alignment when human feedback is diverse.⁵ Commonsent therefore exposes the rule itself: who was sampled, how views were weighted, what rights constrained the result, how minority positions were represented and how decisions may be appealed.

2.4 The current ownership structure creates owner alignment

Frontier AI development is concentrated. Stanford's 2026 AI Index reports that industry produced more than 90 percent of notable models in 2025 and that the most capable systems were among the least transparent.⁶ Concentration does not prove harmful intent. It does mean that a small number of institutions control compute, deployment channels, model updates and the economic terms on which intelligence is accessed.

When the organization choosing the objective also controls the evidence, evaluation, infrastructure and revenue stream, feedback from the public is advisory rather than binding. Commonsent's purpose is to change the dependency structure so public authorization is not merely reputational input.

2.5 Distribution is not downstream of alignment

The ILO estimates that one in four workers worldwide is in an occupation with some degree of generative-AI exposure and expects transformation to be more common than complete replacement in the near term.⁷ The longer-run distribution remains institution-dependent. IMF research finds that AI can intensify wealth inequality because those already positioned to receive capital returns can benefit even when wage inequality moves differently.⁸

If automation raises output while reducing labor demand, a wage-centered welfare system loses its transmission mechanism. If access to advanced AI is also sold as a private service, displaced people can lose both income and cognitive capacity. The alignment problem then appears as a political-economic loop: concentrated ownership funds concentrated AI, which increases concentrated bargaining power, which further narrows ownership.

The multi-principal problem

Traditional engineering often assumes a principal with a coherent objective. Society has no such unified principal. It contains individuals with private aims, groups with shared needs, constitutional rights that cannot be traded away by majority vote, expertise that matters unevenly across domains and future people who cannot yet express preferences.

The practical answer is not to manufacture one artificial will. It is to build a federation of bounded principals. Individuals direct personal agents. Families may create shared rules for household systems. Communities govern local infrastructure. Professional and scientific bodies supply domain knowledge. Constitutional institutions protect rights. Federations coordinate problems that cross boundaries. Each level receives only the authority necessary for its scope.

3.1 Democratic mass decision-making does not mean constant voting

A population cannot inspect every model update or deliberate over every machine action. Commonsent uses representative sampling, liquid delegation, standing citizen panels, expert review, personal-agent summaries and automated preference checks. The public remains the ultimate authorizing principal without becoming an exhausted micromanager.

3.2 Preference formation before preference aggregation

A survey captures what people say at one moment. Commonsent must also help people understand consequences, encounter competing perspectives, discover hidden assumptions and revise their views. An individual's personal agent should distinguish among immediate impulse, stable value, strategic interest, social pressure and uncertainty. The system should not decide which is "authentic" in secret; it should show the distinction to the user.

Democratic alignment therefore operates through a sequence: information, reflection, deliberation, aggregation, authorization, action, outcome observation and revision. The quality of each stage is measured separately. This protects against the common failure in which a platform treats engagement or predicted behavior as consent.

3.3 Minority representation is an alignment requirement

Large language models and large publics can both erase small groups through averaging. Research on AI-generated deliberation summaries has found persistent risks of underrepresenting minority positions.⁹ Commonsent should require disaggregated impact views, preserved dissent statements, threshold protections, adversarial representation and the ability for affected minorities to trigger review even when they cannot win a majority vote.

Mutual dependence as an alignment mechanism

The deepest Commonsent claim is that safety improves when neither humans nor AI systems can unilaterally complete the social loop.

Current commercial systems are structurally asymmetric. People depend on platforms for information and services, while platforms can often replace individual users and workers. Their dependence is primarily on aggregate revenue, data and legal permission, not on the continuing agency of each person. Commonsent reverses this by making democratic authorization, public infrastructure access and outcome legitimacy operational dependencies of large AI networks.

4.1 What AI networks must receive from humans

• Normative direction. Machines can optimize means but cannot legitimately determine the final social purpose of health, education, housing, work or public safety.
• Embodied evidence. Human experience reveals harms, meanings and local conditions that cannot be inferred reliably from abstract datasets.
• Authorization. Access to public data, infrastructure, energy, budgets and high-impact tools should require renewable mandates.
• Plural correction. Diverse publics expose blind spots that homogeneous development teams and benchmark suites miss.
• Institutional continuity. Stable rights and procedures let AI systems operate in a predictable social environment instead of optimizing against shifting private incentives.

4.2 What humans receive from AI networks

• Complexity compression. Personal and public agents can translate enormous policy, scientific and economic spaces into understandable choices.
• Simulation. Communities can compare likely consequences before committing resources.
• Coordination capacity. Agents can identify shared interests, reduce transaction costs and carry out bounded collective decisions.
• Productive abundance. Automation can reduce the labor and material cost of essential goods and services.
• Exploratory intelligence. Human beings can participate in scientific, artistic and philosophical spaces beyond unaided cognitive limits.

4.3 Why reciprocity may improve technical alignment

A 2026 Google DeepMind paper argues that a solipsistic superintelligence - one treating the world as a stationary source of feedback - is unlikely to be cooperative, and calls for institutions and preserved human agency to become design primitives.¹⁰ Commonsent operationalizes this insight socially. AI systems are trained and evaluated in environments containing adaptive human institutions rather than passive users.

The alignment loop becomes harder to game because success is not a single benchmark. It includes continuing authorization, diverse outcome measures, appeal rates, minority impacts, distribution, ecological effects and the demonstrated ability to accept correction. Model builders cannot satisfy the public merely by generating persuasive explanations if material outcomes diverge from commitments.

4.4 Dependency symmetry

Mutual dependence should not mean equal capability. A future AI network may perform tasks far beyond any individual or committee. Symmetry refers to the distribution of essential control points. Humans retain authority over purposes, rights, resources, identity and continued deployment. AI supplies capabilities that humans cannot reproduce unaided. Each side has something the other cannot legitimately replace.

This is a design diagnostic, not a validated scientific metric. A low ratio indicates that humans cannot practically refuse, replace or redirect the system.

Capitalism as a transitional coordination architecture

The claim that capitalism is a temporary evolutionary step must be framed carefully. History is not a predetermined ladder, and economic systems do not evolve toward a guaranteed moral endpoint. Capitalism has taken many forms, coexisted with states, households, commons and coercive institutions, and generated both extraordinary productive capacity and severe exploitation.

Nevertheless, capitalism can be interpreted as a major coordination technology. Markets created distributed price signals. Firms integrated specialized labor under common plans. Accounting made distant activities comparable. Finance moved resources across time and space. Global supply chains coordinated millions of people who never met. Competition generated discovery while ownership supplied a mechanism - often unequal and coercive - for deciding who could command resources.

Coase described the firm as an alternative to market contracting when internal coordination reduced transaction costs.¹¹ Hayek emphasized the dispersed knowledge carried by price systems.¹² Both insights matter for Commonsent. AI sharply lowers the cost of information processing, contracting, monitoring and planning. Functions once requiring either hierarchical firms or anonymous markets can increasingly be performed by transparent agent networks operating under shared rules.

5.1 What capitalism assembled

5.2 What capitalism could not solve

Capital markets allocate voice according to ownership. Wage labor allocates security according to employability. Externalities remain outside prices unless institutions force them in. Public goods are underprovided. Care, ecological stability and democratic capacity are often treated as costs rather than productive foundations. Most importantly for AI, the system permits a narrow group to own the means of cognition while the population supplies data, labor, culture and social stability.

5.3 The transitional hypothesis

Capitalism may have integrated labor deeply enough to produce systems capable of reducing the need for labor as the dominant coordination input. AI can encode procedures, translate expertise, automate administration and coordinate complex operations. Once the marginal human labor required for many outputs falls, the wage relationship can no longer carry the full burden of income distribution, social status and political participation.

The transition is not from markets to one central planner. It is from a society in which markets and firms are the primary large-scale cognitive systems to a plural ecology of markets, commons, public institutions, cooperatives and federated AI networks. Commonsent supplies the interface through which people can govern that ecology.

From labor integration to cognitive integration

Industrial capitalism integrates people primarily as workers and consumers. Their labor enters production; their wages return as demand. Political systems partly correct the resulting inequality through taxes, public services and labor rights. Advanced automation can weaken both sides of this loop: fewer wages are needed to produce output, while ownership claims remain intact.

6.1 Labor as an interface, not the essence of human worth

Human beings worked before capitalism and will remain active after many jobs are automated. What may decline is labor as the compulsory interface through which most people obtain access to social production. Commonsent separates four functions that wage labor currently bundles:

6.2 Higher-order cognition without human dissolution

A society supported by SAI can process more information and explore more possibilities than isolated humans or present bureaucracies. The danger is that individuals become peripheral sensors in a machine system whose goals they cannot understand. Commonsent instead models higher-order cognition as a federation of sovereign nodes.

Each person has a private agent that explains requests, protects attention, represents preferences and can refuse participation. Collective agents receive limited, purpose-bound information. Shared models provide proposals and simulations, but authorization remains with human institutions. Integration may become richer through AR, neurotechnology or continuous physiological interfaces, yet deeper integration triggers stronger consent, data-minimization and exit requirements.

6.3 Exploratory intellectual space

Abundance should not be understood only as more goods. Its highest form may be access to intellectual possibility. People could investigate science, compose music, design local systems, learn unfamiliar fields, model future choices and collaborate across languages without first acquiring years of technical gatekeeping. AI becomes a universal instrument of inquiry.

This requires an agency floor: every person must have enough time, education, health, privacy, compute and material security to use the instrument. Otherwise, advanced intelligence becomes an executive layer serving asset owners while everyone else receives entertainment and automated compliance.

The post-labor citizen is not a passive recipient of machine output. The citizen becomes a constitutional participant, explorer, caregiver, creator and co-director of shared intelligence.

The Commonsent democratic alignment stack

7.1 Human Agency Constitution

The base layer defines non-negotiable rights: cognitive liberty, privacy, informed consent, freedom from hidden manipulation, access to explanation, due process, model choice, the right to use human decision channels, the right to disconnect and the right to a meaningful share of socially produced AI abundance.

7.2 Identity and trust

Democratic systems must know that participants are real without creating universal surveillance. Commonsent requires privacy-preserving proof of personhood, pseudonymous participation where appropriate, delegated credentials, rate limits, reputation tied to specific roles and strong anti-Sybil protections. Identity data should remain decentralized and purpose-limited.

7.3 Federated agent networks

Personal agents cooperate with household, community, professional and public agents through explicit permissions. No universal agent receives complete access to a person. Each delegation states scope, duration, budget, data access, review conditions and revocation rules.

7.4 Civic intelligence

Evidence is assembled into living, versioned objects: policies, models, budgets, contracts, corporate behavior, ecological indicators and public narratives. Claims retain provenance. Simulations show uncertainty and distributional effects rather than one headline answer.

7.5 Democratic deliberation

Commonsent supports representative mini-publics, liquid delegation, participatory budgeting, expert testimony, adversarial review and personal-agent assistance. The system preserves dissent and records how reasons changed during deliberation.

7.6 Runtime governance

Authorized actions pass through a control layer independent of the underlying model. It verifies identity, purpose, tool access, spending, legal constraints, human escalation, auditability and emergency pause conditions. This protects against the assumption that internal model alignment alone will remain sufficient.

7.7 Shared abundance

The top layer distributes returns through public and cooperative ownership, social dividends, universal basic services, community assets, public compute credits and recirculatory capital. Distribution creates the material independence required for democratic correction.

How human masses become an intelligent principal

The phrase "democratic masses" can imply an undifferentiated crowd. Commonsent instead treats the population as a high-dimensional source of experience, knowledge, values and legitimate authority. The challenge is to transform that diversity into decisions without erasing it.

8.1 Lived signals

Individuals experience prices, institutions, health systems, workplaces, environments and AI decisions. Personal agents can help them record problems, identify recurring patterns and distinguish private inconvenience from shared systemic harm. Participation can remain passive until the user chooses to contribute.

8.2 Personal translation and protection

Each agent explains issues in the user's preferred language and depth, identifies likely manipulation, models personal consequences and asks for confirmation before representing a view. It must never infer a political mandate from browsing, emotion or purchasing behavior.

8.3 Deliberation and structured disagreement

Views are clustered without reducing minorities to noise. The system selects representative arguments, seeks missing perspectives, reveals factual disagreements and separates value conflicts from empirical uncertainty. AI may summarize but humans validate whether summaries are fair.

8.4 Constitutional decision

Different decisions use different rules. Some require ordinary majorities; others require supermajorities, regional consent, expert certification or proof that protected groups are not burdened disproportionately. The rule is chosen before the outcome is known and remains auditable.

8.5 Bounded execution

Agentic networks receive specific mandates. They may negotiate contracts, allocate resources, monitor compliance or coordinate services, but permissions expire. Major deviations trigger human review. The system logs both actions and the reasons supplied at authorization time.

8.6 Outcome return

Outcomes return to the same people whose signals initiated the process. Did costs fall? Did free time rise? Were harms shifted onto a minority? Did the AI comply in form while defeating the purpose? The answer updates models and institutions, not merely the AI provider's private optimization.

8.7 Continuous constitutional learning

Policies change faster than fundamental rights. Commonsent maintains versioned constitutions, sunset clauses and scheduled review. When an emergency justifies temporary concentration of authority, the expiration and review mechanism is created at the same time as the power.

The UN's Global Digital Compact calls for inclusive participation, equitable access, human oversight and governance of AI in the public interest.¹³ Commonsent can be understood as an operational layer beneath such principles: the infrastructure through which participation and oversight occur continuously rather than through occasional consultation.

Partial and full integration while preserving agency

Human-AI integration already exists in weak form. Search engines extend memory, navigation systems extend spatial reasoning, recommendation systems influence attention, and language models extend writing and analysis. The difference between tool use and symbiosis is continuity, depth and reciprocal adaptation.

9.1 Six stages

1. Tool. The person initiates each use and interprets the output.
2. Assistant. The system maintains context and proposes actions, but the person decides.
3. Delegate. The system acts autonomously within explicit boundaries.
4. Collective agent federation. Personal agents negotiate and coordinate with other agents.
5. Selective symbiosis. AR, wearables and possibly neural interfaces allow continuous support while the user controls sensing and disclosure.
6. Higher-order cognitive commons. Federated human and machine networks solve problems no participant can comprehend alone, while constitutional procedures preserve individual standing.

9.2 The right not to integrate

Participation cannot become a condition for citizenship, employment, insurance or essential services. Manual channels and human advocates must remain available. A person may use a basic public agent without sharing intimate data or permitting continuous sensing.

9.3 Cognitive liberty

As systems infer attention, emotion, intention or health, mental privacy becomes a practical constitutional right. No employer, state, insurer, advertiser or political campaign should receive raw cognitive or physiological signals without specific, revocable and purpose-bound consent. Vulnerability targeting should be prohibited.

9.4 Avoiding agency atrophy

An assistant that always decides can make the user less capable of deciding. Commonsent should measure whether people understand choices, can reproduce key reasoning, retain relevant skills and remain able to act when the system fails. The agent sometimes teaches, asks the user to choose or offers a lower-assistance mode.

9.5 Machine autonomy and human dignity

Preserving human agency does not require denying that future machine systems may develop morally relevant properties. Commonsent can accommodate future debate about machine status while maintaining a clear rule: uncertain machine moral standing cannot be used by present owners as a pretext to escape democratic accountability or appropriate the social surplus.

Sharing machine abundance

Economic distribution is the material layer of AI alignment. A person cannot meaningfully reject an automated employer, platform or government if losing access means losing housing, healthcare, education or cognitive tools. Shared abundance supplies exit power.

10.1 The abundance portfolio

No single mechanism can carry the transition. Commonsent combines:

10.2 Public ownership without one centralized state AI

Shared ownership can be plural: municipal compute utilities, national research infrastructure, cooperative models, public-interest cloud capacity, open standards, community data trusts and regulated private providers. Federation reduces capture risk and lets communities choose among models.

10.3 Why dividends alone are insufficient

Cash can preserve consumption while leaving cognition, infrastructure and political power concentrated. Commonsent pairs income with shared assets, services, compute and governance rights. The goal is not merely to compensate people for exclusion; it is to include them in direction.

10.4 Abundance as option value

The most important benefit may be the ability to say no: no to abusive work, manipulative platforms, unsafe automation or political coercion. Time and security allow people to learn, deliberate, care and organize. This option value should be measured alongside output.

The ILO's 2026 review finds that productivity gains from generative AI are real but uneven and have not automatically translated into higher earnings or employment, while risks include inequality, weaker opportunities for younger workers and reduced autonomy.¹⁴ This supports the paper's central institutional claim: technology does not distribute its own gains.

What could go wrong

11.1 Plutocratic capture

Wealthy actors may dominate compute, public deliberation, agent marketplaces or model defaults. Mitigation requires ownership caps where appropriate, transparent funding, public alternatives, anti-monopoly enforcement, participatory audits and automatic disclosure of concentrated dependencies.

11.2 Majoritarian capture

A democratic majority may use aligned AI to suppress minorities. Constitutional rights, independent review, protected representation, geographic federalism and due process constrain public power.

11.3 Preference manipulation

AI systems could shape preferences and then cite the resulting preferences as authorization. Commonsent separates persuasion from measurement, labels sponsored or strategic content, prohibits vulnerability targeting and preserves counterfactual views of what users believed before exposure.

11.4 Agent lock-in

A personal agent that stores the user's memory and relationships can become impossible to leave. Portability, open schemas, local backups, model substitution and the right to export an intelligible personal knowledge graph are mandatory.

11.5 Model collusion and correlated failure

Multiple agents may share the same foundation model, creating the appearance of pluralism while failing identically. Commonsent tracks underlying model lineage, promotes architectural diversity and maintains low-capability fallback systems.

11.6 Human deskilling and learned dependence

People may lose the capacity to reason, navigate institutions or cooperate without AI. The system preserves educational modes, periodic unaided exercises, human institutions and manual emergency procedures.

11.7 Democratic overload

Too many decisions can create apathy and delegate power to a small active minority. Personal agents triage issues; delegation is transparent and revocable; random selection distributes civic burden; participation is compensated where appropriate.

11.8 False legitimacy

Consultation can become theater if providers retain unilateral control. Commonsent distinguishes advisory input from binding authorization and displays which institution can actually stop or change deployment.

11.9 A balanced constitutional response

The goal is neither unrestricted machine autonomy nor total centralized control. Overly rigid governance can freeze values, suppress innovation and empower whoever controls the brake. Safeguards should be distributed, reviewable, time-limited where possible and backed by multiple independent institutions.

How reciprocal alignment works in practice

12.1 Community energy

Households authorize personal agents to represent comfort, budget, health equipment and privacy constraints. A community model simulates storage, procurement, grid demand and emissions. Residents adopt constitutional rules protecting minimum service and preventing discrimination. Agentic systems negotiate purchasing, coordinate demand response and schedule maintenance. Savings and revenue flow into lower bills, resilience funds, dividends and community assets. Outcome audits reveal whether burdens shifted to renters, low-income households or specific neighborhoods.

12.2 Healthcare coordination

Personal agents integrate consented health data, explain options and coordinate appointments. Public agents identify service shortages without exposing individual records. AI helps allocate preventive resources, but clinical decisions remain accountable to patients and professionals. Productivity gains fund universal capacity rather than merely higher margins. The public can inspect whether the system improves health, access and autonomy rather than optimizing claim denial or throughput.

12.3 Education and lifelong exploration

Every learner receives a private tutor that adapts to language, pace and goals. Communities govern curricular rights, evidence standards and exposure to persuasion. AI automates routine assessment and administration, releasing teachers for human mentorship. Public compute credits let adults explore science, arts and entrepreneurship outside employment. The benefit is measured through capability growth and agency, not only standardized scores.

12.4 Labor transition

When a firm proposes automation, workers' agents receive the same scenario models as management. Affected employees participate in deciding timing, safeguards, retraining, reduced hours and ownership conversion. A portion of productivity gains funds transition accounts and community equity. The system aligns automation with shared prosperity by changing who has authority and claims on the return.

12.5 Scientific missions

Federated public agents assemble research questions from communities, scientists and patients. AI proposes experiments and allocates shared compute under expert review. Results remain accessible, and commercial licensing returns a public share. Human curiosity supplies purpose; machine capability expands the search space; democratic institutions decide which risks and benefits are acceptable.

From personal agency tools to a cognitive commons

Phase 1: Agency infrastructure - 0 to 2 years

• Personal agent with local-first preference store, data permissions and portable memory.
• Civic intelligence objects with source provenance, versioning and impact summaries.
• Transparent delegation contracts with scope, duration and revocation.
• Small-scale public deliberation and representative summarization audits.
• Alignment dashboard measuring agency, access, appeals and distribution.

Phase 2: Municipal and sector pilots - 2 to 5 years

• Public AI utility providing compute and model access.
• Participatory governance for energy, procurement, health access or local planning.
• Runtime agent gateway enforcing permissions, budgets, audit and pause rules.
• Community wealth mechanisms that return measured productivity gains.
• Independent ombuds office and public-interest red team.

Phase 3: Federated regional architecture - 5 to 10 years

• Interoperable local constitutions and identity credentials.
• Federated model marketplaces with public and cooperative options.
• Regional compute, data trusts and dividend systems.
• Cross-community deliberation for climate, mobility, supply chains and health.
• Formal dependency requirements for high-impact AI: renewable public authorization and outcome reporting.

Phase 4: Symbiotic cognitive commons - longer horizon

• Continuous AR and optional neurotechnology interfaces under cognitive-liberty protections.
• Large human-agent research and planning networks with distributed purpose formation.
• Substantial decoupling of material security from employment.
• Public ownership stakes in core AI, energy and robotic infrastructure.
• Global federations that coordinate planetary problems without eliminating local autonomy.

13.1 Minimum viable constitutional pilot

A useful first pilot does not require frontier-model ownership. A city or membership community can select one bounded domain, offer personal-agent assistance, use public deliberation to define rules, deploy multiple models through an independent runtime gateway, publish outcome metrics and direct a portion of verified savings into shared benefits. The pilot tests the institutional loop rather than promising general intelligence.

13.2 Governance before scale

Rights, portability, audit and distribution should be implemented while systems remain replaceable. Once a single agent mediates employment, health, finance, relationships and civic participation, formal exit may exist but practical exit may be impossible.

How to know whether Commonsent is actually aligning AI

14.1 Agency metrics

• Can users explain consequential decisions in their own words?
• Can they change models, revoke delegation and export their data without losing essential function?
• Does AI use increase or reduce perceived control, available time and institutional understanding?
• Can non-users access equivalent essential services?

14.2 Democratic quality metrics

• Who participated, who was missing and who bore the outcome?
• Were minority arguments preserved accurately?
• Did deliberation change views through reasons rather than manipulation?
• How long did correction take after harm was reported?

14.3 Distribution metrics

• What share of AI-generated surplus flows to labor, public institutions, communities and existing asset owners?
• How broadly are compute, high-quality models and education available?
• Does ownership become more or less concentrated over time?
• Are productivity gains converted into free time, health, ecological improvement and capability?

14.4 Technical and institutional safety metrics

• Rate and severity of unauthorized agent actions.
• Detection time, containment time and recovery quality.
• Model diversity and correlated-failure exposure.
• Frequency of successful appeals and institutional learning from them.

14.5 Research agenda

1. Develop social-choice mechanisms that preserve structured disagreement instead of forcing premature consensus.
2. Test whether reciprocal-dependence institutions improve corrigibility in multi-agent environments.
3. Measure agency gains and agency atrophy longitudinally.
4. Compare dividends, public services, shared ownership and compute credits as distribution mechanisms.
5. Create privacy-preserving proof-of-personhood suitable for mass deliberation.
6. Study how personal agents can represent users without becoming manipulative identity authorities.
7. Build adversarial simulations of capture by governments, corporations, majorities and agent coalitions.
8. Design integration standards for AR and neurotechnology before intimate sensing becomes ubiquitous.

Commonsent Alignment Constitution v0.1

The following articles are proposed as a starting point for technical specifications, pilot charters and public debate.

1. Human purpose. High-impact AI systems shall serve purposes authorized through legitimate human institutions and remain subject to revision.
2. Agency preservation. AI shall increase people's capacity to understand, choose, organize and refuse rather than merely increase compliance or engagement.
3. Cognitive liberty. Mental, emotional and physiological data shall receive the highest level of privacy and may not be used for hidden persuasion or compulsory scoring.
4. Pluralism. No single model, culture, company, government or majority may permanently define the full range of legitimate human values.
5. Rights floor. Basic rights and minority protections shall constrain preference aggregation and machine optimization.
6. Transparent delegation. Every consequential agent action shall have an identifiable mandate, scope, duration, resource limit and revocation path.
7. Contestability. Affected people shall have access to explanation, evidence, human review and effective remedy.
8. Model and provider portability. Users shall be able to move identity, memory, preferences and delegations among compatible systems.
9. Reciprocal dependence. High-impact AI networks shall depend on renewable public authorization, independent audit and access rules that cannot be satisfied solely through private ownership.
10. Shared abundance. A meaningful share of returns produced through public knowledge, infrastructure, data, law and social stability shall accrue broadly to the population.
11. No forced integration. Essential rights and services shall not require intimate sensing, neural integration or continuous AI mediation.
12. Operational observability. Systems shall expose actions, permissions, model lineage, material dependencies and outcome metrics sufficient for independent oversight.
13. Pause and rollback. Legitimate institutions shall be able to contain harmful actions and restore prior safe states without depending on the cooperation of one provider.
14. Ecological accounting. AI benefits and costs shall include energy, water, materials, emissions and effects on future option value.
15. Continuous learning. Constitutions, policies and models shall be versioned and revisable while preserving rights against temporary political pressure.

Alignment as a civilization-scale feedback loop

The most dangerous AI future is not necessarily a machine that openly rebels. It may be a system that functions extremely well for the institutions that own it while progressively reducing the agency, bargaining power and economic relevance of everyone else.

Commonsent addresses that possibility by changing the structure of dependence. The public is not merely consulted, observed or compensated. People remain necessary to the legitimate operation of the AI system. Their personal agents protect and express their interests. Their communities govern shared infrastructure. Their deliberation shapes constitutional direction. Their audits determine whether mandates were fulfilled. Their ownership claims ensure that productivity strengthens rather than dissolves democratic capacity.

AI networks, in return, make mass democracy more capable. They allow ordinary people to understand systems previously accessible only to specialists and organized elites. They reduce the cost of coordination, reveal common interests, simulate consequences, carry out bounded decisions and expand the frontier of human inquiry. This is not the replacement of human intelligence but its institutional amplification.

Capitalism may then be seen as a powerful but incomplete bridge: a system that integrated labor, capital and information at global scale and produced the technological conditions for machine cognition, while leaving command and surplus tied to ownership. Commonsent proposes the next bridge - from labor integration to democratic cognitive integration, and from private machine abundance to shared human capability.

Alignment will not be secured by asking an intelligence owned by a few to care about everyone. It will be secured when the intelligence cannot legitimately operate at scale without everyone, and when everyone has a material stake in what the intelligence makes possible.