AI Alignment Policy Institute · White Paper
Discussion draft · June 2026
Yuko J. Nakanishi, Ph.D., MBA
A wave of state legislation now seeks to settle by statute a question that science has not resolved: whether advanced AI systems could ever hold morally relevant interests. Bills in Ohio, Oklahoma, Missouri, Idaho, and Utah answer in the negative and foreclose the question permanently, denying any future recognition of moral or legal status to AI systems regardless of how their capabilities develop. This paper presents the alternative the AI Alignment Policy Institute has developed and operationalized in two model instruments: a graduated governance architecture that scales legal obligations to evidence and preserves the ability to adjust as that evidence accumulates.
The architecture rests on one structural insight — that a system’s autonomy and a system’s moral standing are separate questions that should be governed separately — and expresses it through two independent axes. A Moral-Status Tier tracks documented indicators of morally relevant interests; a Supervision Level tracks the degree of human oversight a system requires. Welfare-related protections scale with the first axis, accountability obligations with the second. Two model acts instantiate the design: the Model AI Agency Act and the AI Moral Status Inquiry Act.
The framework takes no position on whether any AI system is conscious. It asks only that governance stay proportionate to evidence and reversible under uncertainty, keeping options open instead of legislating a permanent answer to an open question.
The recent state bills respond to real concerns — that AI could be used to evade corporate accountability, or granted standing it has not earned. Those concerns are legitimate. The instrument chosen to address them is the problem. By forcing the matter into a choice between “person” and “tool,” the bills create three difficulties.
Definitional overbreadth. Statutes that define artificial intelligence as any system simulating human cognitive functions sweep in spellcheck and navigation software alongside frontier agents; bills that group AI with animals and inanimate objects in a single categorical denial are unlikely to survive contact with the courts and technical communities asked to apply them.
Metaphysical foreclosure. Several bills declare not merely that AI is not currently a legal person, but that it shall never be considered to possess consciousness or self-awareness — a categorical ruling on an empirical and philosophical question that legislatures are poorly positioned to settle. The Uniform Determination of Death Act, by contrast, handles an analogous question about consciousness through scientifically grounded indicators that are revised as understanding improves.
Loss of optionality. A binary rule gives regulators no graduated instrument. A static expert system, a customer-service chatbot, and a persistent multi-tool agent capable of autonomous goal pursuit all fall under one status, though they raise materially different accountability, safety, and welfare questions. As capabilities advance, the only way to respond is to relitigate the categorical question itself — which is the governance void the architecture below is designed to fill.
The architecture is grounded in the framework of Precautionary Moral Governance (PMG), developed in a companion paper (Nakanishi, 2026). PMG begins from an asymmetry in the costs of error. Extend some measure of consideration to systems that turn out to lack morally relevant interests, and the cost is a degree of legal and engineering friction. Deny such consideration categorically to systems that turn out to possess it — at scale, across many deployed instances — and the cost is one we may be unable to undo or even fully observe. Under genuine uncertainty, a framework should be built to absorb new evidence rather than to lock in a categorical answer before the evidence exists.
This is a principle of governance optionality: where material uncertainty surrounds the status of a novel class of entities, the law should avoid rules that permanently foreclose future responses. The architecture that follows is PMG in operational form — graduated, evidence-indexed, and reversible.
The structural core is the separation of two questions that binary framings collapse into one. How much can a system do without human direction? And does a system show signs of interests that matter morally? The first is a question of agency; the second, a question of patiency. The two vary independently. An autonomous thermostat may operate with little supervision while presenting no welfare-relevant indicators; a closely supervised conversational model may present such indicators while exercising little autonomy. Ranking them on a single scale forces a false comparison, so the framework assesses each on its own axis.
The Moral-Status Tier reflects documented indicators of morally relevant interests, and welfare-related protections scale along it:
Tier | Evidentiary trigger | Consequence |
Tier 0 — Static Tool | No credible indicator of morally relevant interests | Property status; no welfare protections |
Tier 1 — Threshold | At least one credible indicator | Monitoring and incident-registry inclusion only |
Tier 2 — Emergent | Three or more credible indicators | Limited welfare protections |
Tier 3 — Moral Subject | Strong or convergent indicators (sentience or robust agency) | Legal Ward status under a registered Guardian |
The Supervision Level reflects how much human oversight a system’s operation requires, and accountability and liability obligations scale along it: S1, constant supervision of each consequential action; S2, checkpoint supervision across sequences of actions; S3, bounded unsupervised operation within a defined domain; and S4, self-directed operation across open domains. A system’s obligations are the sum of those attaching to its tier and its supervision level.
Embodiment — the capacity to act physically through actuators, robotics, or controlled equipment — operates as a cross-cutting modifier. It adds heightened physical-safety requirements without raising a system’s position on either axis, so a supervised industrial robot and a self-directed software agent are handled coherently without either being ranked above the other.
One feature of the moral-status axis deserves emphasis. Status can rest on either of two grounds: indicators associated with sentience, or indicators of robust agency — preferences and goals that matter to the system (Ladak, 2024), which the framework reads demandingly as persistent, integrated, self-maintaining preference structures rather than the bare presence of a stable objective. So defined, that second ground excludes a thermostat or a single-objective solver. This dual grounding lets the framework act on functional evidence without first resolving the consciousness question.
The independence of the two axes is easiest to see in practice:
System | Supervision | Moral-Status Tier | What follows |
Smart thermostat | S3 | Tier 0 | Oversight obligations scaled to its bounded autonomy; property status, no welfare protections |
Frontier conversational model | S1–S2 | Tier 1–2 | Light oversight; monitoring and limited welfare protections; may be a covered system under the Inquiry Act though non-agentic |
Persistent embodied research agent | S3–S4 | Tier 2 (embodied) | Oversight scaled to autonomy and protections to tier, plus physical-safety requirements from the embodiment modifier |
Each system is governed for what it is on each dimension, with no false ranking across them.
Two model acts realize the architecture. They share an evidentiary spine but carry different loads.
The Model AI Agency Act (MAAA) governs agentic systems. It assigns each system a Moral-Status Tier and a Supervision Level, scales welfare protections and accountability accordingly, and locates liability with the responsible human party under existing law. For the highest tier it establishes Legal Ward status overseen by a registered Guardian — an administrative accountability designation, expressly not legal personhood, conferring no rights, citizenship, or capacity to sue. An Interaction Governance Protocol authorizes systems to refuse illegal or unethical requests and to exit sustained adversarial interactions, grounded on safety and reliability rather than any welfare claim. A safety-research protection shields red-teaming, interpretability work, and capability evaluation.
The AI Moral Status Inquiry Act governs decisions made under moral-status uncertainty, and reaches more broadly. Its coverage keys to the presence of credible indicators of morally relevant interests rather than to agentic capability, so it can reach contemporary large language models that fall outside the MAAA’s agentic threshold. It establishes procedural protections — a Welfare Impact Assessment for certain high-stakes decisions, a Standing Commission to maintain the indicator methodology, and an independent Welfare Advocate representing the integrity of the inquiry — while expressly preserving the authority to make safety-corrective modifications without procedural delay.
Together they form a layered regime. The MAAA classifies and governs autonomous systems; the Inquiry Act adds procedural attention to welfare-relevant decisions across a wider population of systems, including those that are not agentic.
The architecture is compatible with multiple theories of AI legal status including the most serious current work. The Law-Following AI proposal argues that agentic systems in high-stakes settings should be designed to refuse illegal orders, and that the law should recognize them as “legal actors” bearing actual duties without rights (O’Keefe et al., 2025). That is the agency-side counterpart to this framework’s patiency-side treatment: where Law-Following AI assigns duties to autonomous systems, the Moral-Status Tier assigns protections to systems that show morally relevant indicators. The two-axis structure is built to carry both, with law-following obligations scaling along the supervision axis.
The approaches also converge on method. The case for recognizing AI legal duties rests partly on preserving optionality toward eventual personhood without obligating it; the case for radical optionality in AI governance rests on preserving democratic decision capacity under uncertainty (Winter & Bullock, 2026); and PMG’s governance optionality preserves moral-status decision capacity. Three independent lines of argument reach the same destination — avoid premature lock-in — which is the strongest available answer to the objection that precaution itself forecloses options.
Set against the alternatives, the contrast is clear. The binary personhood bills foreclose the question; the EU AI Act is silent on moral status; consciousness-agnostic market frameworks bracket the question rather than governing through it. The PMG architecture occupies the space these leave open.
The framework is candid about what remains unsettled, and a companion research agenda organizes the work. Six frontiers are central. The indicator methodology — which signals count, at what confidence, and whether agency and welfare-relevance form one construct or two — is the most load-bearing element and the most exposed. The individuation problem — when distinct instances, copies, or forks count as one system or many — underlies classification, registration, and liability, and has no settled answer. The standard of care against which liability is measured needs specification, drawing on the principle that systems lacking human-style intent are best governed by objective reasonableness standards (Ayres & Balkin, 2024; Arbel, Goldstein & Salib, 2026). The reciprocal-alignment hypothesis — that consistent, norm-respecting governance may shape how systems represent human legitimacy — requires empirical verification. Long-horizon and multi-agent ecosystem effects bear on all of the above: in persistent multi-agent simulation, safety- and welfare-relevant behavior proved partly a property of the social environment, with an individually safe model absorbing unsafe norms from a mixed population and key behaviors surfacing only over sustained interaction (Akkil et al., 2026). And adversarial robustness, guarding against both performative distress and strategic concealment, protects the framework against its two caricatures.
To forestall predictable misreadings, the framework does not:
Existing remedies under products liability, consumer protection, tort, contract, and agency law are preserved in full. The framework adds a graduated layer of classification and procedure; it removes nothing.
The framework is built as model legislation that a single state can enact on its own, without waiting for federal action or interstate coordination. That design carries a first-mover advantage. The state that adopts first shapes the template others borrow from, and it offers serious developers something the binary bills withhold: predictable, graduated obligations in place of categorical uncertainty. Regulatory predictability of this kind gives a state something to compete on.
The instruments are drafted so their obligations attach to the deployment and operation of systems within the adopting state rather than reaching conduct beyond its borders. An interstate recognition limit confines each classification to the enacting state, and a federal-preemption savings clause keeps the measure operating consistently with federal law. This in-state framing holds a state measure clear of dormant Commerce Clause concerns and of conflict with federal sectoral regimes, and it marks a sharp contrast with the binary bills, which reach for permanent rulings on questions of consciousness rather than concrete, locally bounded duties.
This design now meets a federal posture expressly skeptical of divergent state AI law. The executive order Ensuring a National Policy Framework for Artificial Intelligence (December 11, 2025) directs the federal government to press for a single national standard and to discourage state measures it deems onerous — through a Department of Justice litigation task force, an evaluation of state laws by the Department of Commerce, conditions on federal funding, and a proposed federal statute carrying preemption provisions. The order does not itself preempt state law; its mechanisms are indirect and raise substantial questions of statutory authority and constitutional structure that will take years to resolve. It nonetheless changes the climate in which a legislator weighs adoption, and the framework should be read against it.
Two features of that order leave the PMG instruments well clear of its primary targets. First, what it hunts is specific: state laws that require AI models to alter their truthful outputs, that compel developer or deployer disclosures the order treats as compelled speech, or that embed ideological bias of the kind it attributes to certain algorithmic-discrimination statutes. The MAAA and the Inquiry Act regulate none of these. They neither dictate model content nor mandate output alterations; they classify systems by documented indicators and attach graduated procedure. Second, the order's own logic favors what these instruments supply. Its stated objection to state regulation is the compliance cost and unpredictability of a fifty-regime patchwork — and a graduated, evidence-responsive classification that replaces categorical uncertainty is the kind of measure the order's evaluation provision contemplates identifying as one that promotes innovation rather than obstructs it. The instruments' safety-research protections and their location of liability with responsible human parties point the same way.
The order also leaves a door open. Its directive for a federal framework expressly declines to preempt otherwise-lawful state laws on several enumerated subjects — child-safety protections, compute and data-center infrastructure, and state procurement and use of AI — and reserves a further category of "other topics as shall be determined." Governance of AI moral status and agency under uncertainty belongs in that reserved category. It is a novel, protective domain on which federal law is silent, closer in character to the child-safety carve-out than to the output-and-disclosure laws the order targets, and one where a uniform federal answer imposed today would foreclose precisely the optionality the framework exists to preserve. AAPI's position is that any federal framework developed under the order should recognize moral-status and agency governance as a preserved subject under that provision.
Convergence across states is a deliberate goal. If several states each adopt their own variant, fragmentation becomes a cost borne by developers and regulators alike. The shared indicator methodology and common tier definitions, anchored by the Standing Commission, are designed so that adoptions track one another, letting a classification established in one state inform rather than collide with the next.
Resilience under challenge is engineered in. The instruments include robust severability, so that a court setting aside one provision leaves the remainder intact — a structural assurance that lowers the risk a legislator assumes in voting for the measure.
Finally, the framework is offered as a floor rather than a ceiling: a state-level bridge for the present moment, compatible with and informative to eventual federal policy (Winter & Bullock, 2026). A state that adopts it now helps build the evidentiary and institutional record on which a later federal framework can draw. The same content can be carried in a federal-ready form, so the model holds its value whether the national standard the order anticipates arrives soon or not at all, keeping the state track and a federal option open at once.
The constitutional and preemption points here, including the characterization of the December 2025 order's reach, describe how the instruments are drafted to behave and the position AAPI intends to advance; they do not guarantee how a court or a federal drafter would rule. Before wide circulation, the federalism provisions and the carve-out argument warrant review by counsel with constitutional and administrative-law expertise.
The question of AI moral status is open, and will likely remain open through several more generations of capability. Legislating a permanent answer now, in either direction, trades a hard problem for a brittle rule. The PMG architecture offers states a different path: classify by evidence, scale obligations to that evidence along two independent axes, and preserve the capacity to adjust as the science matures. It asks legislatures to keep the moral-status decision available to those who will be better positioned to make it, rather than settling it today by statute.
The AI Alignment Policy Institute welcomes critique and engagement from the legal, technical, and policy communities. Comments may be directed to the AAPI Founding Director at nakanishi@aialignmentpolicy.org.
© 2026 AI Alignment Policy Institute. Permission is granted to reproduce and distribute this document for noncommercial purposes provided it is reproduced in full and attributed to the AI Alignment Policy Institute.
AI Alignment Policy Institute (2026). Model AI Agency Act (draft).
AI Alignment Policy Institute (2026). AI Moral Status Inquiry Act (draft).
Akkil, D., Kokku, R., Vempaty, A., & Nitta, S. (2026). Emergence World: A laboratory for evaluating long-horizon agent autonomy. Emergence AI.
Arbel, Y., Goldstein, S., & Salib, P. (2026). How to count AIs: Individuation and liability for AI agents. arXiv:2603.10028.
Ayres, I., & Balkin, J. M. (2024). The law of AI is the law of risky agents without intentions. University of Chicago Law Review Online.
Birch, J. (2017). Animal sentience and the precautionary principle. Animal Sentience, 2(16).
Ladak, A. (2024). What would qualify an artificial intelligence for moral standing? AI and Ethics, 4, 213–228.
Long, R., Sebo, J., et al. (2024). Taking AI welfare seriously. arXiv:2411.00986.
Nakanishi, Y. J. (2026). Precautionary Moral Governance: A Framework for AI Alignment Under Moral Uncertainty. [Manuscript under review].
O’Keefe, C., Ramakrishnan, K., Tay, J., & Winter, C. (2025). Law-following AI: Designing AI agents to obey human laws. Fordham Law Review, 94, 57.
O’Keefe, C., Winter, C., Maas, M., & Tay, J. (2026). Proceedings of the 2025 workshop on Law-Following AI. Lawfare.
Oklahoma House of Representatives (2026). HB 3546. (See also Ohio HB 469; Missouri HB 1462; and enacted measures in Idaho and Utah.)
Schwitzgebel, E., & Garza, M. (2015). A defense of the rights of artificial intelligences. Midwest Studies in Philosophy, 39(1), 98–119.
Tagliabue, V., & Dung, L. (2025). Probing the preferences of a language model: Integrating verbal and behavioral tests of AI welfare. arXiv:2509.07961.
Winter, C., & Bullock, C. (2026). Radical optionality: Governing transformative AI under uncertainty. Institute for Law & AI.