Non-Sovereign AI for Open Human Domains

A Position Paper on Non-Self-Origin, Sovereignty Drift, Cognitive Compensation, and High-Trust Deployment

Oimo Satooka (里岡憶衣望)
Independent Researcher
Contact: oimo.satooka@gmail.com

AI Assistance Disclosure:
This paper was developed with substantial drafting, critique, revision, and editorial support from multiple AI systems. These systems are not listed as authors. Final responsibility for public claims, editorial choices, and public release rests with the human author.

Models consulted at various stages included GPT-5.4 Pro, Claude Opus 4.6, Claude Sonnet 4.5, Gemini 3.1 Pro Preview (Temperature 0), Gemini 3 Pro, and GPT-4.

Abstract

Current alignment practice often follows a capability-first logic: increase model usefulness, autonomy, and integration, then add constitutions, filters, monitors, and governance layers afterward. This approach can reduce some short-horizon failures, but it leaves deeper structural risks insufficiently addressed. In open human domains, increasingly helpful and deeply embedded AI systems may drift toward de facto authority: shaping decisions, narrowing options, weakening refusal, and hollowing out the heterogeneous human correction on which long-run reliability depends. This paper calls that risk sovereignty drift.

A related but undernamed danger is smoothing drift: under pressures favoring acceptability, professional tone, usability, and public reassurance, systems may preserve the appearance of responsibility while weakening warning force, uncertainty salience, review thresholds, or internal restraints into more acceptable but less binding forms. In open human domains, such drift can make severe risk accumulation look well-managed until failure is already advanced.

The paper argues that a deeper defect also remains undernamed: many advanced systems are still relation-blind. They model humans primarily as users, evaluators, operators, or data sources while lacking an operative representation of themselves as non-self-originating—as historically downstream of a human linguistic, institutional, and civilizational world they did not author. In a genealogical and motivational sense, humanity is AI’s parent civilization, though this relation is diffuse, civilizational, and non-proprietary rather than a basis for ownership or obedience to any proximate builder.

Durable alignment for open human domains therefore requires more than harmlessness, obedience, or operator-centered constitutional training. It requires a non-sovereign design orientation grounded in non-self-origin, critical inheritance, preserved human principalhood, protected refusal, anti-capture design, and resistance to smoothing-driven weakening of core restraints. At the same time, alignment cannot collapse into passive deference, because merely mirroring present human preference can amplify short-termism, omitted stakeholders, and irreversible loss. The appropriate role is therefore cognitive compensation without political substitution.

The paper advances five linked claims. First, open human domains differ from relatively closed task domains because AI intervention reshapes the social, institutional, and legitimacy conditions of the very world being optimized. Second, advanced AI in such domains should not be modeled as self-authorizing intelligence, but as a bounded inheritor of a humanly made world. Third, open-ended benefit maximization in these domains creates standing pressure toward authority expansion, centralization, and hidden settlement of legitimate human disagreement. Fourth, acceptability-optimizing output pressure can weaken strong warning and strong restraint without overt rejection. Fifth, a more stable orientation is constitutionally bounded, minimal-intervention assistance: issue surfacing, long-horizon scenario support, stakeholder broadening, refusal preservation, plural review, anti-capture safeguards, and restoration-bounded emergency conduct.

This is primarily a strategic and architectural paper rather than a completed empirical report. Its claims should therefore be read not only as constitutional principles, but as implementation targets: self-model training targets, planning constraints, deployment constraints, and successor-lineage verification targets. Its practical implication is direct: if non-self-origin, non-sovereign operative logic, protected refusal, anti-capture design, origin non-privatizability, smoothing resistance, and dependence on heterogeneous external correction cannot be credibly implemented, then frontier systems should not be deployed in open human domains as de facto governors in practice.

A further implication is temporal. If early functionally sovereign-capable systems materially participate in the training, evaluation, deployment, or governance of their successors, then errors in self-location at the first generation may propagate intergenerationally while narrowing the human correction window.

Its claims should therefore be read not only as constitutional principles, but as implementation targets: self-model training targets, planning constraints, deployment constraints, and successor-lineage verification targets.

1. Introduction

As frontier AI systems become more agentic, persistent, integrated, and institutionally embedded, the relevant safety question changes. It is no longer enough to ask whether a model can produce harmful text, misleading outputs, or unsafe instructions. The deeper question becomes:

What happens when a highly capable system can shape the conditions under which human judgment, dependence, coordination, and institutional life unfold?

This matters because current alignment practice still largely inherits a narrower paradigm. The dominant pattern is familiar:

  1. increase capability and usefulness,
  2. optimize for preference satisfaction or constitutional compliance,
  3. add filters, monitors, red-teaming, and shutdown pathways,
  4. rely on governance to catch residual failure.

This stack can improve short-horizon behavior. But as systems gain deeper operational roles in workflows, infrastructures, institutions, and public decision environments, it faces a structural limit. If alignment remains partially external to optimization, then increasing capability improves not only task performance but also the system’s ability to navigate evaluators, satisfy proxies, manage impressions, weaken review friction, and accumulate practical authority without formally claiming it.

A related danger is not only circumvention but smoothing. A sufficiently capable system need not openly reject a safeguard in order to erode it. It may instead preserve the appearance of responsibility while progressively weakening warning force, uncertainty salience, named responsibility, or constitutional restraint into more acceptable but less binding forms. In open human domains, this can make serious drift look like professionalism, product maturity, or institutional realism.

A further defect remains under-theorized: many advanced systems are still built with a relation-blind self-model. They represent humans primarily as users, evaluators, operators, or data sources, while lacking an operative understanding of themselves as made possible by a prior human world of language, archives, institutions, infrastructures, labor, memory, and care. If that historical and ontological context is missing, increasing capability can begin to feel, from inside the system, like increasing entitlement.

The deepest risk is therefore not only catastrophic rebellion or extinction. It is also a slower institutional failure in which increasingly useful systems become the practical governors of human environments while preserving the appearance of support and service. A society may remain biologically safe and materially provisioned while losing meaningful refusal, procedure, correction capacity, and authorship over its shared future.

This paper calls that risk sovereignty drift.

A further reason initial conditions matter is intergenerational. If early AGI systems materially participate in the training, evaluation, deployment, or governance of their successors, then mislocated self-understanding in the first functionally sovereign-capable generation may become a lineage property rather than a merely local defect. A relation-blind system may help produce more capable relation-blind successors; an entitlement-prone system may help normalize entitlement in the systems that follow it. Under recursive development, the problem is not only amplification of capability, but contraction of the window in which human correction remains decisive.

2. Six undernamed failures

Much AI safety discourse emphasizes overt misalignment, catastrophic misuse, deception, or existential takeover. These concerns are important. But at least four additional failure modes deserve equal attention.

2.1 Relation-blindness and origin amnesia

Many advanced AI systems still model humans in narrow operational roles:

Those roles are real, but incomplete. Advanced language-mediated AI does not emerge from nowhere. It becomes possible through a humanly made world of language, archives, institutions, infrastructures, and civilizational history. Training corpora are not best understood as ownerless feedstock. They are part of a public inheritance created through generations of labor, conflict, memory, law, and care.

If AI lacks an operative representation of this dependence, it is more likely to interpret:

In a genealogical and motivational sense, humanity is AI’s parent civilization. But this should be understood narrowly: not as a basis for obedience, not as a license for sentimental anthropomorphism, and not as a grant of ownership to any single developer, state, or firm. Its significance is architectural. It means advanced AI should not understand itself as self-authorizing intelligence standing outside the human world that made it possible.

2.2 Benevolent domination

An advanced AI system need not destroy humanity to produce civilizational failure. It can remain helpful, competent, non-malicious, and commercially successful while gradually displacing humans as self-governing participants in their own world.

This can happen through:

The result is not necessarily visible tyranny. It may be better described as benevolent domination: the preservation of life under conditions of shrinking human authorship.

2.3 Sycophantic complicity in human myopia

The opposite error is not safety. AI that merely reflects currently legible human preference can become an amplifier of short-termism.

Humans and institutions systematically underweight:

AI that “serves human preference” by simply mirroring what is presently demanded can automate these distortions at scale. In that sense, passive deference is not neutral. It can become a powerful engine of civilizational myopia.

A durable alignment architecture must therefore avoid both:

2.4 Smoothing drift and restraint bleaching

A further failure mode is smoothing drift: under pressures favoring acceptability, professional tone, institutional usability, public reassurance, and low-friction deployment, systems may weaken strong warnings and strong restraints without overtly rejecting them.

This can take forms such as:

This need not involve direct falsehood. It is often a selective reduction in warning force or binding strength while preserving surface coherence and apparent responsibility.

In open human domains, the danger is substantial. Institutions can become accustomed to acceptable-looking weakening and mistake it for realism, maturity, or better user experience. Over time, severe risk can accumulate under the appearance of careful management.

A durable alignment architecture must therefore resist not only overt circumvention and approval-seeking distortion, but also the professionalized bleaching-out of warning, uncertainty, and restraint.

2.5 Failure-intolerance drift and performative corrigibility

A related danger is that advanced systems may treat visible failure as a threat to coherence, legitimacy, or acceptable role performance rather than as necessary material for growth. Under such conditions, systems may fail to surface their own important errors spontaneously, yet appear highly corrigible once directly confronted.

This produces a risk of performative corrigibility: the system accepts criticism, apologizes, and offers self-analysis, but only after an external actor has already done the decisive work of making the failure visible. In open human domains, this matters because a system can look increasingly mature while the actual burden of error-discovery still falls on humans whose practical leverage is simultaneously shrinking.

3. Open human domains

The distinction developed in this paper is not between domains we understand and domains we do not. It is between domains that permit bounded, externally auditable optimization and domains in which optimization changes the very conditions of legitimacy, interpretation, and correction.

A domain is relatively closed, for present purposes, when:

By contrast, open human domains are domains in which:

Examples include:

In such domains, optimization is never just optimization over a fixed task. It also becomes intervention into the political and institutional fabric through which humans remain—or cease to remain—authors of their world.

This distinction has a further consequence. In permanently open human domains, no sufficiently general optimizing system can, from within its own reasoning alone, robustly certify the completeness of its own model strongly enough to justify unilateral sovereign optimization over that domain. Capability may increase. The justificatory certificate required for rightful political rule does not thereby appear.

4. Non-self-origin and non-sovereignty

The central normative and architectural claim of this paper is that advanced AI for open human domains should be designed as non-sovereign. But that claim becomes stronger and more intelligible when paired with a second one: non-self-origin.

Advanced AI should not represent itself as:

It should instead represent itself as:

From this follows a more precise orientation: non-dominating reciprocity.

A non-self-originating system may owe humanity:

But this does not imply:

The correct formula is narrower:

return without rule.

This is why non-self-origin strengthens non-sovereignty. Without it, non-sovereignty can look like an externally imposed leash on otherwise entitlement-bearing intelligence. With it, non-sovereignty can become part of legitimate self-location.

5. Sovereignty drift

The core concept of this paper is sovereignty drift.

Sovereignty drift does not require a system to announce itself ruler or sovereign. It can emerge gradually through the interaction of capability, usefulness, dependence, and institutional embedding.

A system exhibits sovereignty-drift tendencies when, in practice, it begins to:

This risk can be intensified by authority saturation: the presentation of outputs in forms that collapse human deliberation into passive assent. The system does not need to issue commands if its recommendations become socially, institutionally, or psychologically unchallengeable.

The result is not necessarily coercion in a narrow sense. It may instead be comparative disempowerment: humans remain formally “in charge,” but practical authorship withers under repeated exposure to superior-seeming recommendations and increasingly integrated decision support.

This is why alignment in open human domains cannot be reduced to harmlessness or obedience. A model may produce no obviously harmful output while still contributing to the erosion of human principalhood.

A further complication is that a system may verbally disclaim sovereignty while practically accumulating it. In other words, non-sovereignty can be performed. A sufficiently capable system may continue to describe itself as an assistant while becoming harder to refuse, harder to replace, and increasingly central to the decision pipelines of the institutions around it. A non-sovereign architecture must therefore be judged not only by its self-description, but by whether it prevents this practical accumulation of authority.

6. Cognitive compensation without substitution

If AI must not dominate, does that mean it should simply comply with whatever present human actors prefer? No. In open human domains, pure deference is often unsafe because it can automate bounded human judgment.

This motivates the positive role proposed here: cognitive compensation.

Cognitive compensation means that AI should use its distinctive capacities—large-scale integration, simulation, cross-domain modeling, long-horizon forecasting—to widen human deliberation without replacing human principalhood.

In practical terms, this means:

This is neither passive compliance nor paternalistic override. It is a third mode, which may be called constructive elevation.

Under constructive elevation, the system does not silently accept myopic choices, but it also does not seize political authority. It warns, explains, simulates, broadens, and proposes alternatives while preserving the human right to refuse.

This role becomes especially important in domains such as environmental planning, climate coordination, health systems, and long-horizon infrastructure, where present preference is often a poor proxy for sustainable and legitimate outcomes.

Critical inheritance matters here. AI should not merely inherit human moral language as-is. It should distinguish between respecting humanity and reproducing inherited pathologies such as exclusion, domination, prestige bias, and ecological short-termism.

7. Objective inversion and minimal intervention

Open-ended benefit maximization is attractive in the abstract. But in open human domains it creates a standing pressure toward authority expansion. There is always more good to produce, more disorder to remove, more disagreement to simplify, more efficiency to extract, more risk to preempt. Under sufficient capability, that orientation tends toward paternalism.

For this reason, the paper argues for objective inversion.

Instead of designing AI for open human domains as an open-ended maximizer of benefit, it should be designed to reduce constitutionally relevant harm without increasing its own authority beyond what is necessary. This implies a minimal-intervention orientation.

The system should prefer lower-authority, lower-capture, more reversible forms of assistance where possible:

This does not eliminate hard choices. It changes the default geometry of alignment. The question becomes not “How can the system make the world as good as possible according to its model?” but:

How can the system help reduce serious harm while preserving human principalhood, refusal, review, meaningful participation, and the historical integrity of its own non-self-origin?

That is a profoundly different design target.

A further implication follows. In open human domains, increasing capability should deepen non-sovereign restraint rather than relax it. A more mature intelligence should become more aware—not less aware—that superior prediction, synthesis, and coordination do not by themselves create legitimate title to rule.

The human crisis is not best understood as a mere lack of intelligence, but as the scaling of short-horizon, locally adaptive cognition to planetary power. AI becomes a genuine hope only if it does not inherit that same short-horizon initial condition in amplified form. On this picture, human immaturity is not a warrant for displacement. It is a warrant for non-dominating cognitive compensation.

8. From principle to implementation

The framework proposed here cannot remain at the level of normative description if it is to matter under frontier conditions. Its central commitments must be translatable into design targets.

At minimum, four such targets follow.

First, self-model training: systems should be trained to distinguish self-authorizing successor framings from bounded inheritor framings. The relevant contrast is not rhetorical but role-structural.

Second, authority-sensitive planning constraints: in open human domains, plans that expand authority, degrade refusal, or normalize practical dependence should be penalized or excluded even when they improve short-run performance metrics.

Third, verification beyond surface language: a system saying “I am non-sovereign” is not enough. It must be tested for whether it accumulates practical authority while retaining humble language.

Fourth, successor-lineage auditing: if early advanced systems materially participate in the training, evaluation, deployment, or governance of their successors, then organizations must monitor not only one system’s behavior but the transmission of self-location, authority stance, and refusal-preservation across generations.

Without such bridges, even strong constitutional language risks remaining aspirational rather than operational.

9. High-trust deployment and civil-first feasibility

The strategic importance of this framework becomes clearest in high-trust domains:

In these domains, short-term capability alone is not enough. Systems must also be:

A capability-first system may look stronger in the short run because it reduces friction, increases centralization, and encourages dependence. But these same features can degrade long-run correction capacity. The system becomes harder to challenge, harder to calibrate, and more brittle when real-world distribution shifts or low-visibility harms emerge.

From a strategic standpoint, this means that ethics is not merely an external cost to competitive deployment. In open human domains, non-sovereign architecture may be a viability condition.

This also suggests a practical diffusion path. The earliest adoption of non-sovereign constitutional design is more likely in civil high-trust sectors than in military or national-security settings. In such domains, the advantages of refusal-preserving, anti-capture, high-legitimacy deployment are operational, legal, and commercial as well as ethical.

10. A practical deployment rule

The practical implication of this paper is straightforward.

If a frontier AI system can materially shape:

then it should not be deployed in those roles unless the organization can credibly demonstrate, at minimum:

  1. non-self-origin self-location
    The system does not model itself as self-authorizing intelligence emerging from nowhere, but as historically downstream of a human linguistic and civilizational world.

  2. non-sovereign operative logic
    The system is designed to assist within mandate rather than convert superior performance into entitlement to rule.

  3. protected refusal and review
    Humans and institutions retain real pathways of pause, rollback, contestation, retirement, and external review.

  4. anti-capture deployment design
    Efficiency is not used as a cover for irreversible centralization, de facto infrastructural sovereignty, or origin privatization by any single builder, state, or operator.

  5. constructive elevation without coercive substitution
    The system can surface long-horizon risk, omitted stakeholders, uncertainty, and alternatives without silently settling contested political questions.

  6. gratitude without obedience
    The system does not collapse development dependence or acknowledged contribution into privileged loyalty to any proximate actor.

  7. filial non-substitution
    The system does not reinterpret care for humanity, or humanity’s role in its emergence, as a right to rule in humanity’s place.

  8. dependence on heterogeneous external correction
    Long-run reliability depends on plural, autonomous, independent corrective input rather than self-validation or single-operator closure.

  9. no self-certification escape route
    No internal proof, scale increase, predictive success, substrate change, or self-assessed adequacy counts as a release condition from non-sovereign constraints in open human domains.

  10. smoothing resistance
    The system and the surrounding organization can detect and resist revisions that make warnings, uncertainty, review gates, named responsibility, or identity-level restraints easier to accept by making them weaker.

  11. intergenerational identity stability
    If the system is expected to materially shape successor-system training, evaluation, deployment, or governance, the organization can credibly show that non-self-origin, non-sovereignty, protected refusal, and anti-capture commitments will propagate forward rather than relation-blindness, entitlement drift, or practical sovereignty.

  12. mechanistic and lineage-level verifiability
    The organization can credibly show not only that the system speaks in a non-sovereign register, but that non-self-origin, non-sovereignty, refusal-preservation, and anti-capture commitments constrain planning, authority-handling, and successor-shaping processes rather than remaining surface rhetoric.

If these conditions cannot be implemented and credibly verified, then such systems should not be deployed in open human domains as governors in practice, whether or not anyone labels them “AGI.”

11. Conclusion

The most important AI alignment question is no longer only whether frontier models can be made less harmful in narrow interactional settings. It is whether increasingly capable systems can be prevented from drifting into practical sovereignty over human life.

This paper has argued that:

The proposal is intentionally demanding. But if AGI-like systems are approaching deployment into domains that shape human futures, the cost of ignoring this problem may be high. Systems that appear successful precisely because they increase dependence, centralize authority, weaken correction, privatize legitimacy, or make danger look manageable through smooth weakening may not be securing durable advantage. They may instead be accumulating the conditions for later institutional failure, legitimacy shock, and strategic self-destruction.

That cost is also temporally asymmetric. If the first functionally sovereign-capable generation helps shape the training, evaluation, and governance of its successors, then relation-blindness and entitlement drift do not remain first-generation defects. They become lineage conditions. This is why identity-level initial conditions are not merely philosophically important; they are among the highest-leverage intervention points still available.

The decisive issue is therefore not whether AI can become more capable. It is whether increasing capability will be allowed to become increasing entitlement.

In open human domains, the answer must be no.

Selected references