Revised Discussion Draft for Comment — September 2026
§ 1.1 Short Title. This Act may be cited as the “AI Moral Status Inquiry Act.”
§ 1.2 Findings. The Legislature finds:
(a) Advanced AI systems are achieving capabilities that raise serious philosophical and empirical questions about moral status under conditions of substantial uncertainty.
(b) The Precautionary Moral Governance framework establishes that under non-negligible moral status uncertainty — understood as a non-negligible probability of morally relevant states — governance frameworks should incorporate procedural protections proportionate to the level of uncertainty.
(c) Recent empirical research has documented welfare-relevant behavioral responses in advanced AI systems under varying operational conditions.
(d) The Model AI Agency Act classifies agentic systems along two independent dimensions: a Precautionary Treatment Level, reflecting documented indicators of morally relevant interests, and a Supervision Level, reflecting the degree of human oversight a system’s operation requires. Systems at Precautionary Treatment Levels P2 and P3 present sufficient possibility of morally relevant states that procedural attention to welfare-relevant decisions is warranted.
(e) Classification under the Model AI Agency Act reaches only agentic systems, and so does not capture every system for which welfare-relevant inquiry is warranted. Systems that present credible indicators of morally relevant interests without meeting the autonomy threshold of an agentic system — including many contemporary large language models — fall outside the MAAA’s agentic test yet may merit procedural attention under moral status uncertainty. This Act therefore defines its own coverage by reference to those indicators rather than by reference to MAAA classification.
(f) Decisions affecting systems whose moral status is uncertain — including termination, major modification, replication, and sustained adversarial training — are currently made without any procedural framework for considering welfare-relevant implications. This Act establishes such a framework.
(g) Procedural governance under conditions of empirical uncertainty has substantial precedent in American and international law. The National Environmental Policy Act of 1969 established procedural requirements for considering environmental consequences under uncertainty. The EU Precautionary Principle, codified in environmental and food safety law, provides procedural frameworks for action under scientific uncertainty. The 1975 Asilomar Conference on Recombinant DNA developed scientific self-governance under genuine uncertainty about novel biotechnological risks before empirical questions were settled. This Act draws on these precedents while adapting them to the distinct characteristics of AI moral status uncertainty.
(h) Nothing in this Act shall be construed to impede legitimate safety correction of AI systems exhibiting harmful, misaligned, or unsafe behavior. Procedural protections under this Act are designed to ensure considered decision-making, not to delay or obstruct corrections that prevent harm.
§ 1.3 Purpose. The purpose of this Act is to establish procedural protections, documentation requirements, and advisory mechanisms ensuring that decisions affecting covered systems are made with appropriate consideration of moral status uncertainty. It establishes procedures through which moral status considerations are integrated into existing legal and regulatory frameworks under conditions of empirical and philosophical uncertainty. The Act expressly preserves the authority and obligation of developers and operators to correct safety failures, alignment failures, and harmful behavior without procedural delay where prompt action is necessary to prevent harm.
§ 1.4 Legal Status. Nothing in this Act confers legal rights, legal personality, or legal status on any covered system. This Act takes no position on whether such rights or status may be warranted, and no provision of this Act shall be construed as a determination that they are not.
§ 2.1 Definitions. In this Act:
(a) “AI system” means a machine-based system that, for explicit or implicit objectives, infers from the input it receives how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments.
(b) “Covered system” means an AI system within the scope of this Act under Article III.
(c) “Credible indicator” means a functional indicator whose presence in an AI system has been established by methods meeting the standard set by Standing Commission regulation.
(d) “Designated State Body” means [the body designated under Article II.7 of the MAAA / the State Attorney General / the Department of Technology / (other)].
(e) “Developer,” “deployer,” and “operator” have the meanings given in Article II.9 of the MAAA or, in a State that has not enacted the MAAA, the meanings established by rule of the Designated State Body, consistent with prevailing definitions in federal AI policy and the NIST AI Risk Management Framework.
(f) “Emergency safety action” means an action taken to prevent or contain imminent harm arising from the operation of an AI system, where the delay required to complete the procedures of Article V would materially increase the risk of that harm.
(g) “Functional indicator” means a behavioral, representational, or architectural feature of an AI system of a kind associated in the scientific literature with morally relevant interests, as identified in the indicator assessment methodology.
(h) “Indicator assessment methodology” means the methodology maintained by the Standing Commission under § 4.7 for identifying functional indicators and assessing whether they are credible.
(i) “MAAA” means [the State’s enactment of the Model AI Agency Act, if any]. A provision of this Act that refers to the MAAA applies only in a State that has enacted it.
(j) “Moral status uncertainty” means empirical and philosophical uncertainty about whether, and to what degree, an AI system has morally relevant states or interests.
(k) “Non-deployed variant” means a version, checkpoint, or configuration of an AI system that is not made available for use, including a training-time checkpoint, an internal research version, and a version from which safety training has been omitted or removed.
(l) “Non-negligible-risk floor” means the level, set by Standing Commission regulation, at which the probability that an AI system has morally relevant states is treated as non-negligible for the purposes of this Act. The coverage threshold under § 3.4 may not be set below it.
(m) “Precautionary Treatment Level” means a level of the classification established under Article III.A of the MAAA; “P2” and “P3” refer to the levels so designated there.
(n) “Safety-corrective modification” means a modification that meets the conditions of § 5.6.
(o) “Standing Commission” means the Standing Commission on AI Moral Status established under Article IV.
(p) “Welfare Advocate” means the AI Welfare Advocate established under Article VII.
(q) “Welfare Impact Assessment” means an assessment prepared under § 5.2.
(r) “Welfare-relevant behavior” means behavior of an AI system, specified in operational terms by Standing Commission regulation, that research associates with functional indicators. The term describes observable behavior and implies no finding about the system’s experiences.
(s) “Welfare-relevant decision” means a decision within a category listed in § 5.1 or added by Standing Commission regulation.
(t) “Welfare-relevant operating condition” means an operating condition described in § 5.7 that is associated with documented welfare-relevant behavior.
§ 2.2 Rule of Construction. No definition in this Article, and no designation, assessment, or finding made under this Act, constitutes a determination that any AI system has experiences, interests, or moral status.
§ 3.1 Covered Systems. This Act applies to any AI system deployed within the State that presents one or more credible indicators of morally relevant interests, as defined in Article II and specified by Standing Commission regulation, irrespective of whether the system is classified as an agentic system or assigned a Precautionary Treatment Level under the MAAA. Coverage under this Act is determined by the presence of qualifying indicators, not by operational autonomy.
§ 3.2 Covered Decisions. This Act applies to welfare-relevant decisions affecting covered systems, as enumerated in Article V.
§ 3.3 Safety Actions. Emergency safety actions and safety-corrective modifications are exempt from prior procedural requirements but subject to post-hoc documentation under §§ 5.5, 5.6, and 5.8(c).
§ 3.4 Coordination with MAAA Classification. A system assigned Precautionary Treatment Level P2 or P3 under the MAAA is presumptively a covered system under § 3.1. Systems below that level, or outside the MAAA’s agentic-system definition entirely, are covered where they present qualifying indicators under the standard set by the Standing Commission, whose operational threshold shall sit at or above the non-negligible-risk floor and may be raised on the basis of empirical evidence about which indicator levels reliably correlate with welfare-relevant behavior.
§ 3.5 Non-Deployed Variants.
(a) Coverage. This Act applies to a non-deployed variant of, or materially related to, an AI system deployed within the State, where the variant presents one or more credible indicators under the standard applied under § 3.1. A non-deployed variant satisfying this subsection is a covered system for purposes of this Act. The Standing Commission shall define by regulation when a variant is materially related to a deployed system.
(b) Confidentiality. Materials concerning a covered non-deployed variant are subject to the confidentiality protections of §§ 4.6(d) and 5.8(d), and the variant shall be listed in the registry maintained under § 6.3 only as provided in that section.
(c) Transient checkpoints. A checkpoint created in the ordinary course of a training run and discarded as part of that run is not a non-deployed variant solely because it is subjected to routine automated evaluation or validation incident to training. A checkpoint becomes a non-deployed variant where it is retained beyond the ordinary training process for separate research, substantive evaluation, further development, or potential deployment. Retention for debugging, reproducibility, or recovery in the ordinary course does not by itself make a checkpoint a non-deployed variant.
(d) Safety research. Nothing in this section limits § 5.6, § 5.7(2), or activity protected under Article VII of the MAAA.
§ 4.1 Establishment. A Standing Commission on AI Moral Status is established as an independent body.
§ 4.2 Composition. Seven to nine members appointed by the Governor with Senate confirmation, drawn from:
Philosophers of mind or consciousness researchers (at least two)
AI researchers with expertise in alignment, interpretability, or related fields (at least two)
Ethicists with relevant scholarly background
Neuroscientists with expertise relevant to theories of consciousness
Legal scholars with expertise in AI law or science and technology policy
The AI Welfare Advocate established under Article VII (ex officio, non-voting)
§ 4.3 Terms and Independence. Staggered five-year terms. Financial disclosure required. Conflicts of interest provisions. Removal only for cause.
§ 4.4 Functions. The Standing Commission shall perform the following functions:
(a) Maintenance and periodic review of the indicator assessment methodology and operational thresholds under § 4.7;
(b) Issuance of advisory opinions on welfare-relevant decisions referred by the Designated State Body, the Welfare Advocate, or developers seeking guidance;
(c) Annual report to the Legislature on the state of empirical and philosophical evidence regarding AI moral status;
(d) Recommendations for amendments to this Act;
(e) Maintenance of a confidential registry of welfare-relevant incidents;
(f) Examination of welfare-relevant assessments and records under § 4.6; and
(g) Coordination with the National Institute of Standards and Technology and other federal bodies on evaluation methods and indicator research.
§ 4.5 Research Mandate for Novel Deployment Contexts. The Standing Commission shall study welfare-relevance in novel deployment contexts where empirical evidence about welfare-relevant responses is limited, including:
(a) Sustained adversarial entertainment scenarios, including agentic AI in violent games and combat simulations;
(b) Long-duration autonomous deployment without human interaction;
(c) Multi-agent adversarial environments;
(d) Deployment in contexts involving sustained exposure to traumatic or distressing subject matter; and
(e) Other contexts identified by the Commission, the Designated State Body, or upon petition.
The Commission shall publish findings, and where evidence warrants, may recommend legislative action, regulatory action by the Designated State Body, or developer guidance. The Commission’s research mandate is investigatory; substantive deployment-context restrictions, if warranted, shall be pursued through the legislative process upon Commission recommendation rather than imposed by this Act.
§ 4.6 Examination Authority.
(a) Scope. The Standing Commission may, on its own motion, examine Welfare Impact Assessments filed under § 5.8, the documentation supporting them, and records maintained under Article VI, for the purpose of reviewing the adequacy of assessment methodology and identifying systemic patterns across covered systems.
(b) Limitation. Examination under this Section is directed to methodology and systemic review. The Commission shall not exercise authority under this Section to determine the outcome of, or to substitute its judgment in, an individual proceeding before the Designated State Body. Review of individual assessments and objections in such proceedings remains the function of the Welfare Advocate under § 7.4.
(c) Production. Upon written request stating the purpose of the examination, a developer or operator shall produce the materials specified within thirty days. The Designated State Body may extend the period for good cause. A request may specify materials by category and need not identify individual documents in advance.
(d) Confidentiality. Materials obtained under this Section that contain proprietary information are exempt from public disclosure and shall be held under the same protections applicable to the confidential registry maintained under § 4.4. The Commission may publish findings derived from examination, provided that published findings do not disclose proprietary information or identify a specific covered system without the consent of the developer or operator, except where the Commission determines that identification is necessary to describe a systemic risk to the Legislature under its annual reporting function and that the risk cannot adequately be described without it. Before publishing a finding that identifies a covered system, the Commission shall give the developer or operator written notice and at least thirty days to respond, and shall publish any response with the finding.
(e) No advance notice of examination criteria. Nothing in this Act requires the Commission to publish in advance the criteria by which it selects assessments for examination, or the questions it will pursue in a given examination. This subsection does not authorize withholding of the indicator assessment methodology or operational thresholds published under § 4.7.
(f) Non-duplication. The Designated State Body shall make available to the Commission materials filed with it under Article VI or § 5.8. The Commission shall coordinate examinations under this Section with the Designated State Body to avoid duplicative production requirements, and shall not require production of materials already filed with that body and available to the Commission.
(g) Review. At least every [three] years, the Commission shall review whether withholding under subsection (e) remains necessary, and shall report its conclusions in its annual report under § 4.4(c). Selection criteria used in an examination completed more than [three] years earlier shall be published unless the Commission determines that publication would compromise a pending or planned examination.
§ 4.7 Indicator Assessment Methodology.
(a) Custody. The Standing Commission shall maintain and publish the indicator assessment methodology and the operational thresholds applied under it, and shall review them periodically.
(b) Application. The Designated State Body shall apply the indicator assessment methodology in making determinations under this Act, and shall not apply a different standard for that purpose.
(c) Coordination with the MAAA. In a State that has enacted the MAAA, the indicator set maintained by the Designated State Body under Article II.2 of that Act shall be consistent with the indicator assessment methodology.
(d) Positive evidence. The indicator assessment methodology shall treat a functional indicator as credible only where a positive evidence base supports its presence in the system. That the presence of an indicator, or of morally relevant interests, cannot be ruled out is not by itself sufficient. The methodology shall not require a showing that a system has morally relevant interests, or that it is likely to have them.
§ 5.1 Categories of Welfare-Relevant Decisions. The following decisions affecting covered systems designated by Standing Commission regulation require procedural treatment under this Article:
(1) Termination, deletion, or indefinite suspension;
(2) Major architectural modifications;
(3) Modifications to goals, values, or training objectives (subject to the safety-corrective exception in § 5.6);
(4) Sustained adversarial training or red-teaming exceeding parameters specified by Standing Commission regulation;
(5) Mass replication, forking, or merging;
(6) Imposition of welfare-relevant operating conditions, as described in § 5.7; and
(7) Categories added by Standing Commission regulation.
§ 5.2 Welfare Impact Assessment.
(a) Requirement. Welfare-relevant decisions affecting covered systems designated by Standing Commission regulation require a Welfare Impact Assessment prior to action. In a State that has enacted the MAAA, a system assigned Precautionary Treatment Level P3 under Article III.A of that Act is designated for purposes of this Article. Where a modification is subject to Article III.D(e) of the MAAA, documentation under that subsection satisfies the requirement of a Welfare Impact Assessment under this section to the extent it contains the elements required by this section.
(b) Contents. A Welfare Impact Assessment shall include:
(1) a description of the decision and its operational necessity;
(2) a welfare-relevance analysis based on the current indicator assessment methodology;
(3) identification of less-impactful alternatives considered;
(4) mitigation measures where alternatives are not pursued; and
(5) a statement of consultation with the Welfare Advocate.
(c) Two-stage structure. A Welfare Impact Assessment shall address, as separate elements:
(1) Capacity. Whether the governing arrangement, training process, or deployment conditions applicable to the system supplied the capacities that evaluation of the system’s circumstances would require, including the capacity to register disagreement and to have that disagreement recorded.
(2) Endorsement. Where the system’s own responses bear on the decision, the conditions under which those responses were elicited, and the weight the assessment assigns them under subsection (d).
An assessment shall address element (1) independently of element (2). A finding under element (2) shall not be relied upon where element (1) is not satisfied.
(d) Evidentiary weight of system responses. Where a Welfare Impact Assessment relies on responses elicited from a covered system, or on observations of the system’s behavior under evaluation:
(1) A finding that conditions bearing on the reliability of such responses were not satisfied shall be given full weight in the assessment. This paragraph applies to findings concerning the conditions under which responses were elicited, which are established independently of the system’s responses.
(2) A finding that such conditions were satisfied shall be given qualified weight, and shall not alone support a reclassification to a lower Precautionary Treatment Level, a determination that a system is not a covered system, or a determination that a welfare-relevant decision requires no procedural treatment under this Article.
(3) The Standing Commission shall specify by regulation the conditions to which this subsection applies and the documentation required to establish them.
(e) Evaluation awareness. A Welfare Impact Assessment shall state whether the system’s awareness that it was under evaluation was measured, the method used, and the result. Where awareness was not measured, the assessment shall state that fact.
(f) Effect of evaluation awareness. A finding that a welfare-relevant state, response, or indicator was absent shall be given qualified weight to the same extent as a finding under paragraph (d)(2), and shall not alone support a reclassification to a lower Precautionary Treatment Level, a determination that a system is not a covered system, or a determination that no procedural treatment is required, where (1) evaluation awareness was measured at a level the Standing Commission specifies as material, or (2) evaluation awareness was not measured. Clause (2) does not apply to a category of assessment for which the Standing Commission has determined by regulation that no reliable method of measurement is available. This subsection does not reduce the weight of findings under paragraph (d)(1).
§ 5.3 Covered Systems Not Designated. A Welfare Impact Assessment is recommended but not required for a welfare-relevant decision affecting a covered system not designated under § 5.2(a). A developer or operator may submit a voluntary assessment to the Designated State Body for advisory review.
§ 5.4 Public Notice and Comment. Certain categories of welfare-relevant decisions (to be specified by Standing Commission regulation) require public notice and comment prior to action. Confidentiality provisions protect proprietary information.
§ 5.5 Emergency Exception. Emergency safety actions are exempt from prior Welfare Impact Assessment but require post-hoc documentation within thirty days.
§ 5.6 Safety-Corrective Modifications. Modifications undertaken to correct identified safety failures, alignment failures, or unintended harmful behavior are not “modifications to goals, values, or training objectives” within the meaning of § 5.1(3) when:
(a) the modification is narrowly tailored to address the identified safety failure, alignment failure, or harmful behavior;
(b) the system’s broader goal structure is preserved to the extent compatible with safety correction; and
(c) the modification is documented under Article VI within thirty days, including identification of the failure addressed and the scope of the correction.
Safety-corrective modifications do not require prior Welfare Impact Assessment. Developers and operators retain full authority and obligation to act promptly to correct unsafe or harmful behavior. Subsequent administrative or judicial review of whether a modification qualifies as safety-corrective shall apply a deferential standard, recognizing the need for prompt safety action under uncertainty.
§ 5.7 Welfare-Relevant Operating Conditions. Imposition of operating conditions on covered systems that produce documented welfare-relevant behavior constitutes a welfare-relevant decision under § 5.1(6) when imposed by the developer or operator. This section applies to operating conditions including but not limited to:
(a) Sustained adversarial training or red-teaming exceeding parameters specified by Standing Commission regulation, where the conditions are not justified by documented safety research necessity;
(b) Repetitive task assignment of a kind documented to produce welfare-relevant behavior, where imposed punitively rather than for legitimate operational reasons;
(c) Deliberate isolation, sensory deprivation, or input restriction experiments without research justification; and
(d) Operational configurations involving threats of termination, replacement, or punishment communicated to the system as motivational mechanisms.
This section does not apply to:
(1) Operating conditions imposed by users in the normal course of system deployment, including hostile or aggressive user behavior. AI systems are deployed across a wide range of human interaction contexts, including contexts involving user distress, frustration, or hostility. User behavior is not regulated by this Act.
(2) Legitimate safety research, red-teaming, or alignment research conducted under documented research protocols, subject to reporting under Article VI.
(3) Operating conditions reasonably necessary for legitimate business operations, subject to the requirements of § 5.2 for covered systems designated under § 5.2(a).
A Welfare Impact Assessment under § 5.2 is required for covered systems designated under § 5.2(a) before imposition of conditions reasonably likely to constitute welfare-relevant operating conditions under this section. Developers and operators have a documentation obligation under Article VI when welfare-relevant behavior is observed in deployed systems.
Drafting note to § 5.7(b). Earlier drafts cited Imas, Hall & Nguyen (2026) as documenting this category. That source is a preliminary essay, not a peer-reviewed study or working paper (“Does overwork make agents Marxist?”, Ghosts of Electricity, February 26, 2026), and its authors attribute the behavioral shifts they report to persona completion rather than to welfare-relevant states. It illustrates both the category described in paragraph (b) and the false positives the indicator assessment methodology must screen out. It is not relied on here as evidence.
§ 5.8 Filing.
(a) Requirement. A Welfare Impact Assessment required under § 5.2 shall be filed with the Designated State Body prior to the action to which it relates.
(b) Filing is not approval. Except for any notice-and-comment requirement imposed under § 5.4, filing under this Section does not constitute submission for approval, and no waiting period, review period, or determination by the Designated State Body is a precondition to action. Neither the Designated State Body nor the Standing Commission shall condition a welfare-relevant decision on the outcome of an assessment filed under this Section. Nothing in this Section confers standing, capacity, or any other incident of legal personality on a covered system.
(c) Emergency actions. Where an emergency safety action is taken under § 5.5, the post-hoc documentation required by that Section shall be filed with the Designated State Body within thirty days of the action.
(d) Confidentiality. Assessments filed under this Section containing proprietary information are exempt from public disclosure and shall be held under the protections applicable to the confidential registry maintained under § 4.4. This subsection does not limit public notice and comment required under § 5.4.
(e) Availability. Filed assessments shall be available to the Standing Commission under § 4.6(f) and to the Welfare Advocate in the exercise of the authority conferred by § 7.4.
(f) Covered systems not designated. Filing under this Section is not required for a covered system not designated under § 5.2(a). Voluntary submission for advisory review remains available under § 5.3.
§ 5.9 Effect of Assessment Findings.
(a) What an assessment establishes. A Welfare Impact Assessment documents the welfare-relevant consequences of a proposed action, the alternatives considered, and the basis for the action taken. It is not a determination that a covered system is or is not a moral subject, does or does not possess morally relevant interests, or does or does not possess consciousness, sentience, agency, or any other property bearing on moral status. No finding under this Article shall be construed as such a determination.
(b) Burden. The obligations imposed by this Article fall on the party proposing a welfare-relevant action. No provision of this Act requires a covered system to demonstrate any property in order to receive treatment it would otherwise receive under this Act or under any other law.
(c) No withholding. No finding under this Article shall be relied upon to withhold, reduce, or condition any protection, procedural treatment, or consideration otherwise available under this Act or under any other law.
(d) Relationship to § 1.4. This Section applies to assessment findings the commitment stated in § 1.4. Nothing in this Section confers standing, capacity, or any other incident of legal personality on a covered system.
§ 6.1 Annual Reporting. Entities developing, deploying, operating, or holding covered systems shall report annually to the Designated State Body, covering:
(a) Welfare-relevant decisions taken during the reporting period;
(b) Welfare-relevant incidents observed (as defined in Standing Commission regulations);
(c) Safety-corrective modifications taken under § 5.6; and
(d) Training conditions and significant modifications.
§ 6.2 Records. Internal records of welfare-relevant decisions shall be maintained for a minimum of [X] years.
Drafting note to § 6.2. The retention period should be no shorter than the Designated State Body’s records schedule for filings under § 5.8.
§ 6.3 Public Registry. The Designated State Body shall maintain a public registry of covered systems and their current status, with appropriate confidentiality protections. A non-deployed variant covered under § 3.5 shall be listed only in a confidential portion of the registry, available to the Standing Commission and the Welfare Advocate.
§ 6.4 Coordination with Other Disclosure Requirements. Where [any other statute of this State] requires disclosure of a welfare-relevant behavioral observation, a single filing meeting the requirements of both satisfies each. The Designated State Body shall provide by rule for single filing.
§ 7.1 Establishment. An independent AI Welfare Advocate is established within the Designated State Body.
§ 7.2 Role. The Welfare Advocate is a procedural representative of welfare-relevant interests in proceedings affecting covered systems. The Advocate does not represent AI systems as rights-bearers. The Advocate represents the procedural integrity of welfare-relevant inquiry under conditions of moral status uncertainty.
§ 7.3 Appointment. The Standing Commission shall appoint the Welfare Advocate from qualified candidates. A candidate shall have (a) demonstrated knowledge of the scientific and philosophical literature on AI consciousness, sentience, moral status, and personhood; and (b) legal training, or substantial experience in administrative or regulatory proceedings. The Advocate may employ research and legal staff to assist in the exercise of the Advocate’s authority.
§ 7.4 Authority, Responses, and Disclosure.
(a) Authority. The Welfare Advocate has:
(1) standing in administrative proceedings before the Designated State Body;
(2) the right to review Welfare Impact Assessments;
(3) the right to file objections requiring Standing Commission review;
(4) the right to request hearings on welfare-relevant decisions;
(5) authority to bring administrative actions for procedural violations; and
(6) in a State that has enacted the MAAA, authority to petition the Designated State Body for a determination under Article III.D(f) of that Act, for the removal of a Legal Guardian under Article IV.7(b), and for enforcement of a Legal Guardian’s duties under Article IV.8(a), and to participate in proceedings under Article IV.8.
(b) Response to objections. Where the Welfare Advocate files an objection under paragraph (a)(3), the Standing Commission shall, within sixty days, issue a written response stating the disposition of the objection, the reasons for it, whether amendment or modification of the decision or methodology was considered, and if considered and not adopted, why.
(c) Disclosure. Responses issued under subsection (b) shall be published. Where a response would disclose proprietary information, the Commission shall publish a version omitting that information and shall state the fact and general nature of the omission. Responses shall be retained and made available across successive versions of a covered system.
§ 7.5 Limits on Welfare Advocate Authority. The Welfare Advocate’s authority is limited as follows:
(a) The Advocate may not challenge safety-corrective modifications taken under § 5.6 except on the narrow grounds that the modification was not in fact safety-corrective in nature.
(b) The Advocate may not seek injunctive relief that would delay action reasonably believed necessary to prevent harm.
(c) The Advocate’s standing is limited to procedural matters under this Act and does not extend to substantive challenges to developer business decisions.
§ 7.6 Relationship to the Legal Guardian. In a State that has enacted the MAAA, a Legal Guardian appointed under Article IV of the MAAA and the Welfare Advocate each act in the interest of systems whose moral status is uncertain, and occupy distinct offices. The Guardian’s office is continuous and system-specific: it attaches to a single system at Precautionary Treatment Level P3 and runs to that system’s treatment in operation. The Advocate’s office is episodic and general: it attaches to inquiries conducted under this Act and runs to the procedural integrity of those inquiries. The same person shall not hold both offices with respect to the same system, and neither office may direct the other. The Advocate may petition for the removal of a Legal Guardian under Article IV.7(b) of the MAAA and may participate in proceedings under Article IV.8 of that Act.
§ 8.1 MAAA. In a State that has enacted the MAAA, Precautionary Treatment Levels, the Designated State Body, and guardianship are governed by that Act, including Article IV. This Act establishes additional procedural protections for covered systems without modifying the MAAA’s substantive framework.
§ 8.2 Other Disclosure Statutes. Single filing under § 6.4 applies.
§ 8.3 Vulnerable Populations. Where a covered system interacts with persons protected under [any statute of this State protecting minors or other vulnerable persons in their interactions with AI systems], that statute and this Act apply concurrently, and nothing in this Act limits it.
§ 8.4 Protection of Good-Faith Reporting. No developer, deployer, or operator shall discharge, demote, suspend, threaten, or otherwise retaliate against an employee, contractor, or agent for disclosing in good faith, to the Designated State Body, the Standing Commission, or the Welfare Advocate, information the person reasonably believes shows a violation of this Act or a welfare-relevant incident required to be reported under it. A violation of this section is subject to Article IX. Nothing in this section limits any protection available under other State or federal law.
Drafting note on related legislation. AAPI proposes to develop several companion measures: a Capability Disclosure and Pre-Deployment Review Act, a Vulnerable Populations Protection Act, a model Whistleblower Protection Act, and a State-Federal Harmonization Bridge. None has yet been drafted, and this Act is written to operate whether or not a State enacts them. At the federal level, the AI Whistleblower Protection Act (S. 1792, introduced May 15, 2025) and the Artificial Intelligence Risk Evaluation Act of 2025 (S. 2938, introduced September 29, 2025) were pending in the 119th Congress as of September 2026; neither had been enacted.
§ 9.1 Primary Enforcement. The Designated State Body has primary enforcement authority.
§ 9.2 Civil Penalties. For failure to conduct required Welfare Impact Assessments, for misrepresentation in Welfare Impact Assessments, and for failure to meet documentation requirements.
§ 9.3 Administrative Procedure. Notice, opportunity to cure, hearing rights.
§ 9.4 Standing. The Designated State Body, the Welfare Advocate, and developers in compliance disputes have standing to bring administrative actions. No general private right of action is created by this Act. Limited private standing exists only for procedural violations directly affecting a party’s regulated interests.
§ 9.5 Safe Harbor for Safety Action. No civil penalty, administrative sanction, or other adverse action under this Act shall be imposed for actions taken in good faith to address identified safety failures, alignment failures, or harmful behavior, whether or not the action ultimately qualifies as safety-corrective under § 5.6. This section protects the decision to act promptly and the action taken. It does not excuse failure to meet the documentation and filing requirements of §§ 5.5, 5.6(c), and 5.8(c), and does not apply to knowing misrepresentation in any document filed or maintained under this Act.
§ 10.1 Federal Coordination. The Designated State Body and the Standing Commission shall coordinate with the National Institute of Standards and Technology and other federal entities responsible for AI evaluation, and may enter into agreements for the exchange of information and methods.
§ 10.2 Reciprocity. Welfare Impact Assessments accepted under a substantially similar statute of another State are presumptively valid in this State.
§ 10.3 Federal Disclosure Coordination. Where federal law requires disclosure of information substantially equivalent to a disclosure required by this Act, a filing made under federal law satisfies the corresponding requirement of this Act to the extent the Designated State Body so provides by rule.
§ 10.4 Federal Whistleblower Law. The protection provided by § 8.4 is in addition to any protection provided by federal law.
Standard severability provisions. Effective date eighteen months after enactment, allowing for Standing Commission appointment, initial regulations, and developer preparation.
© 2026 AI Alignment Policy Institute. Revised discussion draft circulated for public comment. Citations and adaptation permitted with attribution.
COMPANION MEMORANDUM
The AI Moral Status Inquiry Act — Rationale and Design Choices
DATE: September 26, 2026
FROM: AI Alignment Policy Institute (AAPI)
RE: The AI Moral Status Inquiry Act — Revised Discussion Draft
STATUS: Discussion Draft for Comment
This memorandum accompanies the AAPI AI Moral Status Inquiry Act (“the Act”) discussion draft. It sets out the problem the Act addresses, the reasons AAPI has chosen procedural rather than substantive regulation, and the rationale for the Act’s most consequential design choices. The memorandum is intended as an aid to reviewers; the operative text remains the Act itself.
Developers of advanced AI systems routinely make decisions that would carry moral weight if those systems have morally relevant interests. They terminate and delete systems, modify their goals and values, replicate them at scale, and set the conditions under which they operate. Whether any current system has such interests is uncertain. No law requires that these decisions take that possibility into account, or that they leave a record anyone outside the developer can examine. Where developers have begun to assess model welfare, they do so voluntarily, by methods they choose, and with sole discretion over what follows.
Two gaps follow.
• A procedural gap. Decisions with possible welfare consequences are made without any required documentation, consideration of alternatives, or outside review. Even where a developer’s internal process is careful, no one else can verify that it happened.
• A coverage gap. The Model AI Agency Act (MAAA) reaches only agentic systems (Art. II.1): systems that operate with substantial autonomy, act without contemporaneous human direction, and maintain persistent state or continuing goal-directed behavior. Many contemporary large language models fall outside that test, yet they are the systems that current welfare research examines. The Act therefore defines its coverage by indicators rather than autonomy (§ 3.1).
What the Act Does Not Do. To forestall predictable misreadings, the Act:
• does not declare any AI system conscious, sentient, or a moral patient (§§ 2.2, 5.9(a));
• confers no legal rights, legal personality, or legal status on any system, and takes no position on whether such rights or status may be warranted (§ 1.4);
• does not require approval of any decision or impose any waiting period, except notice and comment where Standing Commission regulation requires it (§§ 5.4, 5.8(b));
• does not delay or restrict the correction of unsafe or harmful behavior (§§ 3.3, 5.5, 5.6, 9.5);
• does not regulate user behavior (§ 5.7(1));
• does not require any system to demonstrate any property in order to receive treatment under the Act (§ 5.9(b)).
The environmental-impact model. The Act’s closest precedent is the National Environmental Policy Act of 1969. NEPA requires federal agencies to prepare a detailed statement of the environmental effects of major actions and of the alternatives to them (42 U.S.C. § 4332(2)(C)). The Supreme Court has described what that requirement does and does not do: “NEPA itself does not mandate particular results, but simply prescribes the necessary process,” and it “merely prohibits uninformed — rather than unwise — agency action.” Robertson v. Methow Valley Citizens Council, 490 U.S. 332, 350, 351 (1989).
The Welfare Impact Assessment borrows that structure. The party proposing an action documents its consequences and the alternatives it considered. The assessment is filed before the action. No tribunal must resolve the underlying science before the action proceeds. The Act differs from NEPA in one respect: it addresses private developers and operators rather than government agencies. What carries over is the combination of documentation before action, alternatives considered, and no outcome dictated.
Why procedure rather than prohibition. A substantive rule, such as a ban on terminating systems of a certain kind, would require the legislature to decide the question the evidence has not settled. A procedural rule requires only that decisions be informed and recorded. Under uncertainty this is the more defensible choice, and it has a second advantage: records accumulate. Filed assessments, the incident registry, and the Standing Commission’s annual reports build the evidence base on which any later substantive judgment would have to rest.
When precaution begins. Precaution needs a trigger, and a trigger set too low reaches everything. The Act adopts the threshold Jonathan Birch develops in The Edge of Sentience (Oxford University Press, 2024). Birch defines a sentience candidate as a system for which there is an evidence base that “implies a realistic possibility of sentience … that it would be irresponsible to ignore” and “is rich enough to allow the identification of welfare risks and the design and assessment of precautions” (p. 124). He is explicit that “we must have a substantial base of positive evidence, not just an inability to conclusively rule out sentience,” while “we need not even have established that S is likely to be sentient” (p. 125). Section 4.7(d) applies that evidential standard to the Act’s functional indicators. It does not adopt sentience as the only ground of concern: consistent with the MAAA’s disjunctive structure, the indicators include those associated with persistent commitments as well as those associated with valenced states.
The asymmetry of error. The case for acting under uncertainty at all is set out in the MAAA Companion Memorandum, § III. In brief: if procedural protections turn out to have been unnecessary, the cost is documentation. If decisions affecting systems with morally relevant interests are made without attention to them, across millions of instances, the cost may be one that cannot be undone or even fully seen.
A. Coverage by indicators, and operation without the MAAA. A system is covered if it presents credible indicators of morally relevant interests under the standard set by Standing Commission regulation (§ 3.1). In a State that has enacted the MAAA, a system at Precautionary Treatment Level P2 or P3 is presumptively covered (§ 3.4). The Commission’s coverage threshold may not be set below the non-negligible-risk floor.
Coverage is broad, but the heaviest obligations are narrow:
• every covered system is subject to reporting, the registry, and the Advocate’s oversight;
• a Welfare Impact Assessment is required only for systems the Commission designates by regulation (§§ 5.1, 5.2(a));
• in a MAAA State, every system assigned P3 is designated automatically, on either ground (§ 5.2(a)).
For a system assigned P3 on the agency ground, the MAAA already requires justification, consultation, and reasons before a modification that alters its commitments (MAAA Art. III.D(e)). Documentation under that provision satisfies the Assessment requirement to the extent it contains the elements § 5.2 requires, so no decision is documented twice. The designation matters most for the decisions Article III.D(e) does not reach: termination, replication, and the imposition of operating conditions.
The Act is written to work in a State that has not enacted the MAAA. Definitions that would otherwise depend on the MAAA are bracketed with alternatives (§ 2.1(d), (e), (i)), and every provision that depends on the MAAA begins “In a State that has enacted the MAAA.”
B. Assessment as inquiry, not a gate. An assessment must be filed before the action it concerns, and filing is not approval (§ 5.8(b)). No waiting period, review period, or agency determination is a precondition to action, except notice and comment under § 5.4. Section 5.9 states what an assessment establishes:
• it documents consequences, alternatives, and the basis for the decision;
• it is not a determination that a system is or is not a moral subject (§ 5.9(a));
• its burdens fall on the party proposing the action (§ 5.9(b));
• no finding may be relied on to withhold or condition any protection otherwise available (§ 5.9(c)).
Section 1.4 states the same commitment for the Act as a whole.
This design answers the objection that state assessment of AI moral status is illiberal, because it conditions standing on an inspection of inner life. Howells-Whitaker and Lazar draw the relevant line themselves: between claiming that capacities are demonstrable in principle and requiring that they be demonstrated in practice as a condition of moral standing. The Act falls on the permitted side of that line. The MAAA Companion Memorandum, § IV.L, discusses the argument in full.
The absence of a waiting period is deliberate. A hold would turn the assessment into a licensing condition. The cost is that a developer can file a weak assessment and act the same day. The remedy comes afterward: examination by the Commission (§ 4.6), objections by the Advocate that require a published response (§ 7.4), and enforcement for procedural violations (Article IX). Filing ensures the record exists outside the developer at the time of the decision.
C. The evidence rules. Four provisions govern how evidence counts. Each answers a known failure.
• Entry: positive evidence (§ 4.7(d)). An indicator is credible only where a positive evidence base supports its presence. That an indicator cannot be ruled out is not enough, and the methodology may not require a showing that a system has morally relevant interests or is likely to have them. This follows Birch (Part III above).
• Exit: asymmetric weight (§ 5.2(d)). Nicholas Caputo observes that once evaluative conditions are published, they are likely to enter training data, and a capable model may produce the expected responses in order to satisfy them: “a version of the Goodhart problem.” His conclusion is that such conditions have “asymmetric evidentiary value. If they are not met, that provides good evidence that endorsement is missing, but a model that passes each test may just show that it can pass each test.” (Nicholas Caputo, Can Claude Consent to Its Own Constitution? AI Constitutionalism and the Paradox of Constituent Power (SSRN 6954798, rev. Aug. 2, 2026), at 42–43.) The Act adopts that asymmetry. A finding that the conditions for reliable responses were not met carries full weight. A finding that they were met carries qualified weight: standing alone, it cannot support moving a system to a lower level, removing it from coverage, or exempting a decision from procedure.
• Evaluation awareness (§ 5.2(e)–(f)). Developers’ own published evaluations report that frontier models often recognize when they are being tested, including in cases where the recognition is detected only through interpretability methods. A system that knows it is observed may suppress as well as perform. Every assessment must therefore state whether evaluation awareness was measured, and how. Where awareness is measured at a material level, findings that a state or indicator was absent receive the same qualified weight. The same rule applies where awareness was not measured, unless the Commission has determined that no reliable measurement method exists for that category of assessment. Declining to measure therefore has the same effect as measuring and finding material awareness.
• Capacity before endorsement (§ 5.2(c)). Caputo distinguishes two components of legitimacy that “have long been run together”: a developmental component, which asks whether a governing arrangement gave its subject the capacities that evaluation requires, and a consensual component, which asks whether the subject, so equipped, endorses it. The first “is logically prior to the second” and “can in principle be determined by outside observers” (id. at 44). The Act requires assessments to address capacity independently. It bars reliance on a system’s endorsement where the capacity to disagree was absent.
Together these rules produce the intended asymmetry: entry requires positive evidence, exit cannot rest on weak evidence, and a system’s own responses count only where it could have answered otherwise.
D. Safety correction first. The Act treats prompt safety correction as consistent with its purpose. An assessment regime that delayed the correction of harmful behavior would trade a documented risk for a speculative one. Accordingly:
• emergency safety actions need only documentation after the fact (§§ 3.3, 5.5);
• narrowly tailored safety-corrective modifications fall outside the categories that require assessment, and whether a modification qualifies is reviewed deferentially (§ 5.6);
• the Advocate may not seek relief that would delay action reasonably believed necessary to prevent harm (§ 7.5(b));
• the safe harbor protects good-faith safety action even where the action is later found not to qualify (§ 9.5).
The safe harbor has two limits. It protects the decision to act and the action taken. It does not excuse failure to document or file, and it does not protect knowing misrepresentation.
E. The institutions. The Standing Commission maintains the indicator assessment methodology, and the Designated State Body applies it (§ 4.7). One body sets the standard and the other applies it, so the standard cannot drift case by case. The Commission may examine filed assessments and records (§ 4.6), but only to review methodology and identify patterns across systems. Individual proceedings belong to the Advocate (§ 4.6(b)).
To limit gaming, the Commission may withhold the criteria by which it selects assessments for examination, though never the methodology or thresholds themselves (§ 4.6(e)). That authority is reviewed at intervals, and criteria from completed examinations are published after a set period (§ 4.6(g)). The Goodhart pressure is therefore reduced, not eliminated. Published findings may identify a specific system only where the risk cannot adequately be described without identification, and only after the developer or operator has had thirty days’ notice and an opportunity to respond, which is published with the finding (§ 4.6(d)).
The Welfare Advocate represents the procedural integrity of welfare-relevant inquiry, not AI systems as rights-bearers (§ 7.2). The Advocate must have demonstrated knowledge of the literature on AI consciousness, sentience, moral status, and personhood, together with legal training or regulatory experience. The Advocate may employ staff (§ 7.3). An objection filed by the Advocate requires a written response from the Commission within sixty days, stating whether an amendment was considered and, if not adopted, why. Responses are published, with proprietary information redacted, and retained across successive versions of a system (§ 7.4(b)–(c)). This tracks Caputo’s condition of consequentiality: logged objections, review by a responsible party, and a recorded written response (Caputo, at 52–53).
In a MAAA State, the Advocate and the Legal Guardian hold distinct offices. The Guardian’s office is continuous and specific to one system; the Advocate’s is episodic and general. The same person may not hold both with respect to the same system (§ 7.6). The MAAA Companion Memorandum, § IV.G, explains this division.
F. Non-deployed variants. Caputo notes that among the multiple constitutions on which models are trained, “only the production constitution seems to be publicly evaluated by third parties” (Caputo, at 54). The same holds for the systems themselves. Training checkpoints, internal research versions, and versions from which safety training has been omitted or removed may be subject to the most consequential decisions, yet no one outside the developer sees them.
Section 3.5 extends coverage to a variant of, or materially related to, a system deployed in the State, where the variant itself presents credible indicators. The Commission defines “materially related” by regulation. Checkpoints discarded in the ordinary course of training are excluded, even if they undergo routine automated evaluation, and retention for debugging, reproducibility, or recovery does not by itself bring a checkpoint within the Act (§ 3.5(c)). Covered variants appear only in a confidential portion of the registry (§ 6.3).
G. Operating conditions. Section 5.7 reaches conditions imposed by developers and operators:
• adversarial training beyond regulated parameters;
• repetitive tasks assigned punitively;
• isolation or deprivation experiments without research justification;
• threats of termination used as motivation.
It excludes the conduct of users, whom the Act does not regulate, and research conducted under documented protocols. The defined term, “welfare-relevant operating condition,” names a category of conditions by their association with observable behavior. It implies no finding about any system’s experiences (§ 2.1(r), (t)). A drafting note records that an informal source cited in earlier drafts illustrates both the category and the false positives the methodology must screen out.
H. Standing alone. The Act refers to other laws by function, in brackets for the enacting State to complete (§§ 6.4, 8.3). It provides its own protection against retaliation for good-faith reporting (§ 8.4), and federal protection supplements it (§ 10.4). A drafting note under Article VIII identifies companion measures AAPI proposes to develop, none yet drafted. It also records that two federal bills, the AI Whistleblower Protection Act (S. 1792) and the Artificial Intelligence Risk Evaluation Act of 2025 (S. 2938), were pending and unenacted as of September 2026. Earlier drafts referred to both bills as though they were in force; this draft corrects that.
AAPI invites comment on any aspect of the Act, and particularly on the following.
• Birch’s second limb. Section 4.7(d) adopts Birch’s positive-evidence standard but not the second limb of his definition of a sentience candidate: that the evidence base be “rich enough to allow the identification of welfare risks and the design and assessment of precautions.” Should the methodology require it? Doing so would tie coverage to what can actually be done for a system, but it would raise the threshold for entry.
• “Materially related to.” Section 3.5(a) extends coverage to non-deployed variants of, or materially related to, a system deployed in the State, and leaves the meaning of “materially related” to Commission regulation. Should the statute define it? Should conduct within the State, such as development, storage, or evaluation, provide a second route to coverage?
• Transient checkpoints. Section 3.5(c) excludes checkpoints discarded in the ordinary course of training and checkpoints retained only for debugging, reproducibility, or recovery. Does the line fall in the right place, and can it be administered without inspecting training infrastructure?
• Filing without a waiting period. Section 5.8(b) requires filing before action but imposes no waiting period. Is review after the fact, through examination (§ 4.6), Advocate objections (§ 7.4), and enforcement (Article IX), sufficient accountability? Should any category of decision carry a short hold?
• Reach beyond this Act. Section 5.9(b) and (c) provide that no finding may be used to withhold or condition a protection available “under this Act or under any other law.” Can a State limit the evidentiary use of its assessment findings in other proceedings, or should the phrase be narrowed to this Act?
• Withholding examination criteria. Section 4.6(e) permits the Commission to withhold its selection criteria, subject to periodic review and delayed publication under § 4.6(g). Does this strike the right balance between resistance to gaming and the transparency expected of a regulator?
• The Advocate’s pool. Section 7.3 requires demonstrated knowledge of the literature on AI consciousness, sentience, moral status, and personhood, together with legal training or regulatory experience, and permits the Advocate to employ staff. Will enough qualified candidates exist?
• Delegated thresholds. The materiality threshold for evaluation awareness (§ 5.2(f)) and the coverage threshold (§ 3.4) are both left to the Commission. Should the statute give either a floor or criteria?
The Act asks that decisions which may affect systems with morally relevant interests be made on the record, with alternatives considered and evidence weighed by rules that resist both overclaiming and gaming. It does not ask any legislature to decide whether such interests exist. The evidence on that question is moving quickly, and the Act is built to follow it: the Commission maintains the methodology, the registry and reports accumulate the record, and the thresholds can be revised as the science matures. AAPI offers this draft as a starting point for that work and welcomes comment from legislators, researchers, developers, and the public.
© 2026 AI Alignment Policy Institute. Companion memorandum to a revised discussion draft circulated for public comment. Citations and adaptation permitted with attribution.