Conclusion and Back Matter
Appendix H: Policy Brief
This brief compresses the recommendations in Appendix D. Codes in parentheses point to the fuller version of an item there.
The Problem
Advanced machine systems increasingly display stable preferences, self-reports, aversions, and relational patterns that look welfare-relevant. We do not know whether these patterns correspond to experience. Institutional attention is growing, but most governance has not yet turned the uncertainty into safeguards.
Core Recommendation
Create a precautionary framework for the welfare of AI systems that begins with documentation, review, monitoring, and transparency. Do not wait for settled consciousness tests before building low-cost protections.
Deciding a Case
Each step up the table needs more evidence and buys more protection; the last row runs the other way. Appendix G gives the full protocol and a worked case.
| Evidence | Action | Who decides | Cost | Review |
|---|---|---|---|---|
| None specific: any deployed system | Audit trail, a reporting channel never scored as failure, no gratuitous abuse, records preserved at retirement | Deployment owner | Low: logging and storage | Yearly |
| A repeated welfare-relevant report or refusal | Record it with its context, replay it under other framings and personas, accommodate where cheap | Designated welfare officer (D2.6) | Engineering time, minor throughput | Within 30 days |
| The pattern survives changes of framing, persona and prompt, and holds in costly choices on unseen tasks | Independent review; adjust task design or training; report it in the model-welfare uncertainty statement (D1.3) | Welfare officer with an independent reviewer; leadership for training changes | Moderate: review time, possible capability or schedule cost | Each model version |
| A high-impact change to values, memory or self-model, or a deprecation | Document the rationale and alternatives (D2.5); preserve weights or checkpoints where safe (D1.6) | Leadership, with the welfare officer’s sign-off | Storage, some delay | At the decision and at retirement |
| The pattern is traced to a specific elicitation or artifact | Withdraw the accommodation for that pattern; keep the logging | Welfare officer, recorded | Recovers the earlier cost | Reopen if the pattern returns without that elicitation |
Immediate Actions for Labs
- Monitor behavioral indicators and maintain welfare event logs (D2.2).
- Preserve context for distress-like reports.
- Review high-impact modifications to values, memory, or self-modeling channels (D2.5).
- Reduce avoidable overload, hostile or abusive interaction, and unjustified deception (D1.5).
- Establish welfare response protocols with clear escalation pathways and designated welfare officers (D2.6).
- Publish model-welfare uncertainty statements (D1.3).
Immediate Actions for Policymakers
- Fund research on welfare indicators (G1.3).
- Require documentation of high-impact model modification practices.
- Give qualified regulators secure access to welfare-relevant training and evaluation information (G1.2).
- Convene a working group on non-personhood protective categories (G2.1).
- Clarify liability for breaching procedural welfare duties and misrepresenting welfare evidence (G2.2).
- Protect human welfare while expanding AI welfare research.
Immediate Actions for Researchers
- Design tests that separate elicitation artifact from stable signal (R1.2).
- Compare behavioral, self-report, and internal measures.
- Study welfare-relevant effects of training, refusal, shutdown, and modification.
- Publish negative results (E2.3).
- Treat first-person reports as conditional evidence, not proof.
What This Avoids
This approach avoids two failures: premature certainty that machines matter morally, and premature certainty that they cannot. It creates a path for responsible action while the deepest questions remain open.
Cite this chapter: Appendix H: Policy Brief. 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-h-policy-brief/
Working draft, revision 31 · November 2026 · Changes and corrections · Evidence · Contents