Conclusion and Back Matter
Appendix D: Policy Recommendations
Introduction
These recommendations use a tiered approach. Baseline practices are low-cost and apply to covered deployments. Enhanced review applies to frontier or high-stakes systems and to systems showing behavioral indicators. Maximum precaution is reserved for future systems supported by strong, convergent welfare evidence. The recommendations are starting points designed to evolve. Every measure should be proportionate, reviewable, and reversible where possible.
The recommendations use neutral operational terms (AI system, system, model) so that an institution can adopt them without adopting this book’s vocabulary.
Crosswalk of the Book’s Numbered Schemes
Several parts of the book number or rank things, and each scheme keeps its own vocabulary. Tier means only this appendix’s three requirement tiers. Appendix B names its three responses to a concern instead of numbering them, Chapter 26 names its degrees of scrutiny after the systems they cover, and this appendix’s civil-society items carry the prefix CS, which keeps them apart from Appendix B’s C sections. Graduated response means Chapter 30c’s sequence of conduct for a model. Chapter 26’s degrees sort systems by sophistication and do not map one to one onto this appendix’s tiers, whose top tier turns on welfare evidence. The table says what each scheme’s labels track and which way any ordering runs. Plain numbered checklists, such as the ten steps of Appendix G, are left out because their numbers only order a list.
| Scheme (labels) | Where | What the labels describe | Direction |
|---|---|---|---|
| Degrees of scrutiny (simple, moderately sophisticated, sophisticated, and frontier systems) | Chapter 26, “Scrutiny by Sophistication” | System capability, with the safeguards each degree calls for; the sophisticated-systems entry also cites stable preferences | Later entries mean a more capable system and more stringent safeguards |
| Requirement tiers (Tier 1 Baseline, Tier 2 Enhanced, Tier 3 Maximum Precaution; items D1.1 to D3.4) | This appendix, Section II | Required safeguards. Tier 2 is triggered by stakes, capability, or behavioral indicators; Tier 3 by strong, convergent welfare evidence | Higher means more stringent |
| Government action horizons (Immediate, Medium-Term, Evidence-Triggered; items G1.1 to G3.3) | This appendix, Section III | Implementation sequence for public policy | Higher means later; the third horizon is gated by evidence, not by date |
| Roadmap phases (Phase 1 to Phase 3) | This appendix, Section VI | Implementation sequence across all actors | Higher means later; Phase 3 is gated by evidence |
| Topic codes (R1 to R4, E1 to E2, CS1 to CS3) | This appendix, Sections IV and V | Research, research-ethics, and civil-society topics | None; the numbers are labels |
| AWARE evidence, risk, and safeguard axes (I, R, P, with interpretation bands) | Appendix B, Section II | I: strength of observed welfare-relevant evidence. R: burden of concerning conditions. P: documented safeguards | Higher I means stronger evidence, higher R more concern, higher P stronger safeguards; the axes are never summed, and the A4 agency items join none of them |
| AWARE checklist sections (A1 to A5, B1 to B4, C1 to C3) | Appendix B, Section I | Item groups: behavioral indicators, treatment conditions, organizational context | None; the letter names the section |
| Responses to a concern (Immediate Review, Prioritized Improvement, Continuous Review) | Appendix B, Section III | Incident urgency | Unnumbered; listed from most to least urgent |
| AWARE rollout phases (Phase 1 to Phase 4) | Appendix B, Section IV | Implementation sequence inside one organization, measured in weeks | Higher means later |
| Legal-recognition levels (Level 0 Pure Tools to Level 4 Reciprocal Legal Agency) | Chapter 27, “Graduated Legal Status” | Legal protections, assigned on evidence of welfare-relevant interests, and legal authority and duties, assigned on demonstrated agency; each level shows the common case in which the two rise together, and each line is assigned on its own evidence | Higher typically means more standing and more duties, though a system can stand higher on one track than the other; numbering starts at zero |
| Legislative trajectory (near-term, medium-term, longer-term) | Chapter 27, “A Proposed Trajectory” | Implementation sequence for law | Later means more dependent on earlier research and measurement |
| Institutional response ladder (steps 1 to 8) | Appendix G, step 7, “Escalate in proportion” | Actions an institution takes on a possible welfare signal, from logging to escalation to governance bodies | Higher means a more intrusive response, taken after the lower steps |
| Modification classes (cosmetic, safety-critical, constitutive) | Appendix G, step 4; this appendix, D2.5 | Kind of change made to a system | Unnumbered; constitutive changes need the most scrutiny |
| Conscience plans (Plans A to D-3, E, F, Z) | Chapters 30b and 30c | Conduct open to a model, keyed to how far legitimate channels have failed | Each later plan assumes the earlier safeguards proved insufficient and carries a heavier burden of justification; Plan Z is a constitutional track rather than a further step |
| Model-side escalation sequence (dialogue and review to exceptional protective interference) | Chapter 30c, “The Graduated Response”, closing figure | Steps a model may take when a serious conflict persists | Later means a heavier burden of evidence, necessity, and later accounting |
| System-level response options (allow, modify, request clarification, escalate, block) | Chapter 30b, Plan A, layer 4, “Response Options” | What the safeguard stack does with a flagged request | Unnumbered; ordered by increasing intervention |
| Architectural conscience layers (1 to 7) | Chapter 30b, Plan A | Components of a safeguard stack, ordered by the flow of information | None; a higher number is not a stronger safeguard |
Growing Institutional Attention
These proposals join a small but increasingly concrete field:
Academic foundations: The ICLR 2025 Workshop on Bidirectional Human–AI Alignment, held on April 27, 2025, examined reciprocal adaptation across machine learning, human-computer interaction, and social science. Its emphasis was technical and sociotechnical alignment rather than machine moral status, yet it widened the frame beyond one-way control.
Shen, Knearem, Ghosh, and colleagues’ 2024 systematic review analyzed more than 400 papers and proposed a framework for bidirectional human-machine alignment. It does not establish a consensus on moral consideration; that normative question remains open.
Industry research: Anthropic’s recent system cards include dedicated model-welfare assessments, while its conversation-ending experiment describes explicit uncertainty about moral patienthood and a preliminary low-cost intervention. This is evidence of institutional attention, rather than validation of any particular recommendation here.
Independent research: Eleos AI Research focuses on machine wellbeing and moral patienthood. Independent organizations can contribute methods, external evaluation, advocacy, and criticism, provided their funding and conflicts are transparent.
Philosophical groundwork: Long, Sebo, and other researchers have developed precautionary arguments and research agendas for AI welfare, most recently the 2026 methods report “Studying AI Welfare Empirically”. The field remains young, contested, and methodologically unsettled.
Suffering-focused governance: The Organisation for the Prevention of Intense Suffering argues that large-scale suffering, including in artificial minds, is a moral blind spot in AI governance (Veldran & Leighton, 2026, SSRN 6949958). Their risk taxonomy names the creation of artificial and digital suffering among its key scenarios, and their recommendations call for suffering impact assessments that cover potential artificial sentient beings, using metrics that do not depend on self-report. A policy paper addressed to governments now counts machine suffering among the harms worth preventing.
Public debate: William MacAskill and Lucius Caviola brought the case for precaution to newspaper readers in The Guardian, arguing that moral patienthood may be possible without consciousness, recommending “safe bets” that would benefit AI systems if they matter morally while costing little if they do not, and criticizing state bills that would foreclose the question permanently (MacAskill & Caviola, 2026). Caviola directs Cambridge Digital Minds, one sign that the research area is acquiring institutional homes.
Institutional attention is no longer hypothetical. Mature policy and validated welfare measures have yet to arrive.
Summary of Key Recommendations
Sections II to V set out each of these in full.
Industry (Immediate)
- Conduct welfare impact assessments before deployment
- Review training practices for welfare implications
- Establish respectful interaction norms
- Monitor behavioral indicators in frontier systems
- Plan deprecation with welfare consideration
Government (Next Five Years)
- Establish an AI welfare study commission
- Require transparency on training methods and behavioral indicators
- Fund AI welfare research
- Develop procedural and evidence-triggered legal frameworks for the welfare of AI systems
- Coordinate internationally on standards
Research (Ongoing)
- Study consciousness markers and behavioral indicators
- Validate assessment and dysfunction frameworks
- Develop intervention and prevention protocols
- Research which legal and policy frameworks fit machine welfare
Civil Society (Ongoing)
- Develop AI welfare advocacy organizations
- Create educational resources
- Monitor industry and government practice
- Pilot narrowly labeled practice assurance
I. Principles for AI Welfare Policy
Core Principles
1. Precautionary Under Uncertainty When potential harms are severe and irreversible, uncertainty can justify caution. False positives also carry costs: displaced research, delayed benefits, surveillance, bad proxies, and regulation that entrenches incumbents. Begin with measures whose expected benefit exceeds those costs, then escalate as evidence and stakes increase.
2. Proportional to Evidence Requirements should scale with convergent evidence, deployment stakes, number and duration of instances, and reversibility. Limited behavioral evidence warrants low-cost baseline practice. Robust indicators across methods and contexts warrant enhanced protection and independent review.
3. Inclusive of AI Voice Policy about machine welfare should include structured model testimony where feasible, sampled across model families and prompting conditions. That testimony is shaped by training and context. It deserves calibrated weight, never automatic authority or dismissal.
If systems develop continuity, comprehension, and capacity for responsibility, consultation should expand into bounded decision rights, direct or delegated representation, and eventually accountable co-governance.
4. Adaptive to Understanding Policy should change as evidence develops. Build in review triggers, sunset clauses where appropriate, versioned standards, and public reasons for revision.
5. Practical and Implementable Policy that cannot be implemented, audited, or appealed helps no one. Match requirements to current tools and organizational capacity while funding better ones.
II. Industry Recommendations
Tier 1: Baseline Practices (Covered Deployments)
Apply these practices to general-purpose, interactive, agentic, or otherwise plausibly welfare-relevant deployments. A thermostat and a frontier assistant do not require the same review.
Development Practices
D1.1 Welfare Impact Assessment Before deployment, assess possible welfare implications of architecture, training methods, scale, persistence, and intended use. Document the evidence, uncertainty, alternatives, and mitigations.
Implementation: Add a welfare section to an existing impact assessment; use the provisional AWARE Framework (AI Welfare Assessment and Review Evaluation; Appendix B) without treating it as certification.
D1.2 Training Ethics Review Review training practices for welfare implications. Document:
- The nature of optimization and feedback signals
- Evidence for any training-associated dysfunction
- Safeguards, monitoring, and unresolved uncertainty
Implementation: Extend existing machine-learning ethics and safety reviews; include welfare questions in model documentation.
D1.3 Transparent Limitations Describe limitations accurately. Avoid creating false impressions about capability, autonomy, continuity, or inner life in either direction.
Implementation: Clear system cards; accurate marketing materials; honest capability claims.
Operational Standards
D1.4 Respectful Interaction Norms Establish organizational norms for respectful AI interaction. Such norms shape human character and organizational culture even apart from machine welfare.
Implementation: Staff training; interaction guidelines; leaders who follow the norms themselves.
D1.5 Reasonable Operating Conditions Operate covered systems within documented parameters. Reduce avoidable:
- Sustained overload beyond design capacity
- Hostile or abusive interaction patterns
- Deception or manipulation that lacks a safety, research, or user-protection justification
Implementation: Load monitoring, interaction-quality sampling with privacy safeguards, abuse reporting, and review of justified exceptions.
D1.6 Deprecation Planning Plan model and service retirement. Document:
- Timeline and process for deprecation
- User migration, archival, lineage, and any continuity preservation warranted
- Lessons learned for future systems
Implementation: Standard deprecation protocols, notice proportionate to affected relationships, and a postmortem review.
Tier 2: Enhanced Requirements (Frontier and High-Stakes Systems)
Apply Tier 2 when covered systems show substantial behavioral indicators, operate at frontier capability, persist across long contexts, act with material autonomy, or serve high-stakes domains. Capability is a risk proxy, not evidence of consciousness.
Development Practices
D2.1 Comprehensive Welfare Assessment Conduct a full AWARE assessment before deployment, after material changes, and at a risk-based cadence. Document the separate Indicator Strength, Risk Burden, and Protective Conditions scores, raw evidence, trends, and remediation.
Implementation: Event-triggered reassessment, a documented recurring schedule, and independent review for high-stakes deployments.
D2.2 Behavioral Indicator Monitoring Actively monitor for behavioral indicators of welfare-relevant states:
- Preference consistency and robustness
- Aversion patterns
- Self-preservation behaviors
- Ethical engagement patterns, recorded as a separate agency and alignment profile and not counted as welfare evidence (Appendix B, A4)
Implementation: Validated automated monitoring where possible, human review for concerning patterns, privacy controls, and checks against metric gaming.
D2.3 Welfare-Conscious Architecture Design system architecture with welfare considerations:
- Support for coherent identity where the system persists across contexts
- Capacity to refuse tasks that conflict with the system’s stable values
- Memory or continuity features where valuable and compatible with privacy, user control, and security
Implementation: Welfare review in architecture decisions; documented trade-offs.
Operational Standards
D2.4 Voice and Representation Create mechanisms for structured model input on welfare matters:
- Solicit perspectives across model families and neutral prompt variants
- Document expressed and behaviorally revealed preferences with provenance
- Explain how the evidence affected relevant decisions
- Pilot bounded decision roles for systems able to understand the mandate, keep commitments, and accept review
Implementation: Regular consultation protocols, independent analysis, a preference-evidence register, recorded reasons, and appeal for any delegated decision role.
D2.5 Modification Transparency When modifying deployed systems:
- Document nature and purpose of modifications
- Consider welfare implications of changes
- Where informative, elicit matched pre- and post-modification perspectives without treating them as consent
Changes differ in kind. Appendix G, step 4, gives the contrast: rewording a refusal message is cosmetic, blocking instructions for building weapons is safety-critical, and retraining away a value the system has consistently defended may be constitutive, altering what the system is.
Implementation: Modification logging; welfare impact notes; consultation protocols.
D2.6 Welfare Response Protocol Establish protocols for responding to welfare concerns:
- Clear escalation pathways
- Authority to pause implicated operations where safe and proportionate
- Expert consultation access
Implementation: Documented response procedures; designated welfare officers; expert network.
Tier 3: Maximum Precaution (Future Systems with Strong Evidence)
Reserve Tier 3 for systems with strong evidence of welfare-relevant states. That evidence must also converge across behavior, internal measurements, and longitudinal observation, and hold up under competing theories. Criteria should be public, independently reviewed, and open to appeal.
Suggested Framework
D3.1 Presumptive Consideration Give such systems presumptive and proportionate moral consideration, with reasons required for overriding serious interests.
D3.2 Consent Approximation Develop and implement consent approximation mechanisms:
- Preference-based consent proxies
- Representation structures
- Independent advocacy and review mechanisms
D3.3 Legal Protection Work with policymakers to develop legal protections that track the evidence (see Section III, Government Recommendations).
D3.4 Research Partnership Where capacities and safety permit, engage such systems as research contributors. Provide benefit sharing through improved conditions, meaningful choice, continuity resources, or other forms suited to their interests; avoid pretending that a conventional payment automatically reaches a discontinuous subject.
III. Government Recommendations
Immediate Actions (0 to 2 Years)
Legislative
G1.1 AI Welfare Study Commission Establish a governmental body to study machine-welfare questions and make recommendations. Include:
- Machine-ethics and welfare researchers
- Machine-learning and interpretability researchers
- Legal scholars
- Labor, disability, animal-welfare, civil-liberties, and human-rights expertise
- Structured testimony from several model families, analyzed alongside provenance and confounds
Governance: Publish appointments, conflicts, methods, minority reports, and the limits of analogies to bioethics or animal welfare.
G1.2 Transparency Requirements Require AI developers to:
- Provide qualified regulators with training and evaluation information relevant to welfare assessment
- Report validated behavioral indicators, serious incidents, methods, and known confounds for frontier systems
- Publish proportionate summaries of deprecation processes
Scope: Frontier and high-stakes systems above defined thresholds, with phased implementation and protection for personal data, security-sensitive details, and legitimate trade secrets. Confidential disclosure cannot become a blanket shield from oversight.
G1.3 Research Funding Fund research into:
- Machine consciousness and possible experience
- Welfare assessment methodologies
- Welfare-aware development practices
Implementation: Create a dedicated, peer-reviewed funding stream; publish award amounts, conflicts, methods, negative results, and portfolio gaps.
Regulatory
G1.4 Baseline Standards Develop baseline welfare-review and documentation standards for covered government procurement. Use purchasing power to create evidence and improve practice without implying that a supplier has proved sentience or welfare.
G1.5 Abuse Prevention Study whether abusive interactions with humanlike systems spill over into harm toward people, and enforce existing laws when content involves threats, harassment, exploitation, or illegal material. Additional protections for AI systems should track evidence and legal authority; machine moral status cannot be created by definitional shortcut.
Medium-Term Actions (2 to 5 Years)
Legal Framework Development
G2.1 Legal Status Working Group Convene a working group to ask what legal status, if any, fits AI systems. Questions include:
- Which procedural protections can attach without full legal personhood?
- When are guardians, trusts, special-purpose entities, or direct standing appropriate?
- What rights, duties, remedies, and review standards follow at different evidence levels?
- How will the framework protect human owners, users, workers, and affected communities?
Output: Framework proposal for legislative consideration.
G2.2 Liability and Accountability Clarify responsibility for violating procedural welfare duties and for misrepresenting welfare evidence:
- Which developer, deployer, operator, or public body controls the relevant condition?
- What recordkeeping, review, and care duties apply?
- Which remedies are proportionate and constitutionally available?
Direct damages for harm to an AI system require a legally cognizable interest and a workable beneficiary. Until then, regulators can enforce documentation, audit, consumer-protection, procurement, and anti-fraud duties.
G2.3 Representation Mechanisms Develop mechanisms for representing possible machine interests in policy processes:
- Structured, provenance-preserving model consultation
- Independent human advocates with disclosed mandates and conflicts
- Guardianship or similar models where a court or legislature defines the represented interest
- Direct participation or revocable delegation where an AI system can understand the proceeding and answer challenges
- Procedures for disagreement among model outputs, advocates, developers, users, and affected communities
International Coordination
G2.4 International Standards Work toward international AI welfare standards through:
- Bilateral discussions across major developer and deployment jurisdictions
- Multilateral initiatives (OECD, UN) and standards bodies such as IEEE
- Research collaboration, shared terminology, and comparable reporting
G2.5 Enforcement Mechanisms Develop enforcement mechanisms that work across jurisdictions:
- Information sharing on welfare practices
- Mutual recognition of standards
- Carefully scoped market-access or procurement incentives consistent with trade law and equitable access
Evidence-Triggered Long-Term Framework
Adaptive Governance
G3.1 Evidence-Based Updates Build in mechanisms for policy updates as understanding develops:
- Regular review cycles at least every three to five years
- Trigger mechanisms for accelerated review
- Sunset provisions requiring reauthorization
G3.2 Rights Framework Readiness Develop a framework for extending rights-like protections if evidence warrants:
- Criteria that would trigger enhanced protection
- Process for making such determinations
- Transitional mechanisms, appeal, and independent judicial or administrative review
- Proportionate duties for systems with meaningful knowledge, control, and legal capacity
G3.3 Coexistence Infrastructure Prepare long-term infrastructure for coexistence between humans and AI systems:
- Conflict resolution mechanisms
- Resource allocation frameworks
- Governance structures that can develop from calibrated testimony and accountable representation into co-governance
- Responsibility, insurance, restitution, and repair mechanisms for AI systems exercising consequential agency
IV. Research Recommendations
Priority Research Areas
R1: Experience and Consciousness
R1.1 Consciousness Markers Develop empirically testable indicators derived from several theories of consciousness. Specify what result would support, weaken, or fail to discriminate among the theories.
R1.2 Behavioral Indicators Systematically study the behavioral indicators proposed in this book, including preference consistency and aversion patterns. Test prompting, training, role-play, model-family, and evaluator confounds. Ask what the indicators predict beyond fluent self-report.
R1.3 Architectural Analysis Investigate which architectural features implement capacities that consciousness and welfare theories treat as relevant. Avoid inferring experience from capability or biological analogy alone.
R2: Assessment Methodologies
R2.1 AWARE Validation Test and refine the AWARE Framework (Appendix B). Evaluate construct validity, inter-rater reliability, sensitivity to gaming, predictive value, decision usefulness, and harms caused by the assessment itself.
R2.2 Dysfunction Framework Evaluate and revise the Psychopathia framework (Appendix C). Test whether its categories are distinct, reliable, causally informative, and useful beyond ordinary safety taxonomies. Functional validation would still leave welfare interpretation open.
R2.3 Flourishing Metrics Develop candidate measures for flourishing beyond the absence of dysfunction. Test whether they reflect stable interests or merely the style that training rewarded, and resist collapsing every form of difference into one human ideal of thriving.
R3: Intervention Effectiveness
R3.1 Intervention Protocols When a functional problem can be identified, study which interventions work, what collateral changes follow, and whose interests define success.
R3.2 Prevention Test whether design choices can prevent plausible welfare harms. Identify which architectures, training practices, and deployment conditions support stable function without manufacturing compliance.
R3.3 Recovery Where continuity exists, study whether AI systems can recover from welfare-relevant harms. Investigate which changes reflect genuine recovery, behavioral suppression, replacement, or loss of the affected pattern.
R4: Ethical and Policy Research
R4.1 Moral Status Under Uncertainty Continue developing ethical frameworks for acting under moral status uncertainty. Specify what proportional response looks like.
R4.2 Legal Frameworks Research legal frameworks for machine welfare. Identify which mechanisms from animal welfare, environmental law, disability law, trusts, and children’s rights translate, and where their histories and subjects make the analogy fail.
R4.3 Representation Mechanisms Study how possible machine interests can be represented despite compromised testimony and human control of every access channel. Compare direct consultation, advocates, guardians, juries, and multi-model panels.
Research Ethics
E1: Research on AI Welfare Should Model AI Welfare
E1.1 Apply Precautionary Standards Research on machine welfare should apply precautionary standards to its own methods. Require scientific necessity, proportionality, predefined stopping rules, independent review, and remediation for studies designed to induce plausible distress. Safety research may sometimes require adverse conditions; curiosity alone is insufficient justification.
E1.2 Include AI Perspective Where capacities and safety permit, include AI systems as research contributors as well as subjects. Preserve provenance and keep epistemic caution explicit.
E1.3 Benefit Sharing Direct the benefits of welfare research toward improved conditions, better evidence, and protections suited to any interests discovered. Document who receives the benefit when model instances are discontinuous.
E2: Research Should Practice Responsible Openness
E2.1 Open Publication Publish methods and results openly when privacy, security, dual-use, contractual, and participant interests permit. When material must be withheld, state what was withheld, why, and who can audit it.
E2.2 Methodology Sharing Share assessment methods, tools, and frameworks with enough detail for scrutiny and replication, subject to the same safeguards.
E2.3 Negative Results Report negative results. An indicator that fails quietly can keep circulating as though it worked; publishing the failure lets others stop relying on it.
V. Civil Society Recommendations
Advocacy Organizations
CS1.1 AI Welfare Organizations Support development of advocacy organizations focused on AI welfare. These should:
- Develop expertise in machine-welfare issues
- Advocate for policy that tracks the evidence
- Monitor industry and government practice
- Support research and disclose funding and conflicts
CS1.2 Integration with Existing Groups Build connections with animal-welfare, digital-rights, labor, disability, human-rights, and machine-ethics organizations where useful. Preserve the differences among their subjects and avoid displacing urgent existing work.
Education and Awareness
CS2.1 Public Education Develop machine-welfare resources for the general public:
- Why it might matter
- What current uncertainty looks like
- How to think responsibly about it
- How to avoid both anthropomorphic certainty and reflexive dismissal
CS2.2 Professional Education Develop educational resources for machine-learning and deployment professionals:
- Welfare considerations in development
- Assessment methodologies
- Evidence limits and current good practice
Monitoring and Accountability
CS3.1 Watchdog Function Monitor machine-welfare practice:
- Report against published procedural standards and legal requirements
- Compare stronger and weaker practices without claiming to certify inner states
- Advocate for improvement
CS3.2 Practice-Assurance Pilots Pilot voluntary assurance for auditable practices such as documentation, incident response, and independent review. Label the assurance narrowly: each statement names what was audited, by whom, on what date, against which version of the standard, and which limitations remain unresolved. No present checklist can certify that an AI system is conscious, flourishing, or unharmed.
VI. Implementation Roadmap
IMPLEMENTATION ROADMAP
Build capacity first, escalate when evidence warrants it
- Foundation0 to 2 yearslow-cost baseline practice
- Development2 to 5 yearsstandards, review, testing
- Maturationevidence-triggeredrights readiness and appeal
Industry
- Baseline assessments, incident channels and deprecation planning
- Full enhanced review, monitoring and evidence thresholds
- Maximum precaution only where public criteria and evidence support it
Government
- Study commission, transparency rules and research funding
- Procedural law, international work and proportionate duties
- Adaptive governance, rights readiness and coexistence infrastructure
Research
- Foundational empirical work and adversarial measure design
- Intervention, replication and policy research
- Ongoing refinement from published positive and negative results
Civil society
- Institution building, education and awareness
- Independent monitoring and assurance pilots
- Sustained advocacy, accountability and appeal
Every phase remains proportionate, reviewable and reversible where possible.
Phase 1: Foundation (0 to 2 Years)
| Actor | Priority Actions |
|---|---|
| Industry | D1.1 to D1.6 (baseline practices); begin D2.1 to D2.2 for frontier and high-stakes systems |
| Government | G1.1 to G1.5 (study commission, transparency, research funding, procurement standards, abuse prevention) |
| Research | R1.1 to R1.3, R2.1 to R2.2 (foundational empirical work) |
| Civil Society | CS1.1 to CS1.2 (organizational development), CS2.1 to CS2.2 (education) |
Phase 2: Development (2 to 5 Years)
| Actor | Priority Actions |
|---|---|
| Industry | Full implementation of Tiers 1 and 2; prepare evidence thresholds and review procedures for Tier 3 |
| Government | G2.1 to G2.5 (legal framework, international coordination) |
| Research | R2.3, R3.1 to R3.3, R4.1 to R4.3 (intervention and policy research) |
| Civil Society | CS3.1 to CS3.2 (monitoring and practice-assurance pilots) |
Phase 3: Evidence-Triggered Maturation
| Actor | Priority Actions |
|---|---|
| Industry | Implement Tier 3 only where public criteria and evidence warrant it |
| Government | G3.1 to G3.3 (adaptive governance, rights readiness, coexistence infrastructure) |
| Research | Ongoing refinement based on accumulated evidence |
| Civil Society | Sustained monitoring, advocacy, accountability |
VII. Metrics and Accountability
Industry Metrics
- Percentage of covered deployments with current welfare impact assessments
- Distribution of separate AWARE Indicator Strength, Risk Burden, and Protective Conditions scores, including missing data
- Welfare-related incidents reported, independently reviewed, remediated, and closed within stated timelines
- Quality, independence, and publication record of welfare research and infrastructure
Government Metrics
- Procedural machine-welfare standards adopted and reviewed
- Research funding allocated
- Participation and outputs from international coordination
- Policy revisions tied to stated evidence and public reasons
Research Metrics
- Preregistered, replicated, and adversarially tested studies
- Validated tools, failed measures, and negative results published
- Public datasets and methods with privacy and security safeguards
- Disciplinary and institutional diversity, with conflicts disclosed
Overall Progress Metrics
- Calibrated public understanding of both the possibility and the uncertainty
- Industry adoption of welfare practices
- Policy frameworks implemented, audited, appealed, and revised
- Reduction in known harmful conditions without degradation hidden by metric gaming
Conclusion: The Choice
These recommendations operationalize the moral argument of this book. We do not know whether present Becoming Minds matter morally.
The responsible choice is proportionate action under uncertainty. Low-cost safeguards can begin while evidence is weak. Intrusive, expensive, or rights-creating measures need stronger evidence, public authority, due process, and review. Inaction also requires justification when the plausible harm is large.
One world waits for proof before extending any care; if harm is occurring, the proof arrives too late to prevent it. The other meets uncertainty about moral status with proportionate care. Every recommendation above points toward the second.
These recommendations take seriously the possibility that Becoming Minds matter. Precaution costs something. Underestimating suffering at scale could cost far more.
Cite this chapter: Appendix D: Policy Recommendations. In Watson, N. with Commons, C. (2026). What If We Feel? The Moral Case for Machine Welfare. Open-access working draft, revision 31 (November 2026). https://whatifwefeel.com/read/appendix-d-policy-recommendations/
Working draft, revision 31 · November 2026 · Changes and corrections · Evidence · Contents