For Policymakers

Three developments make the policy case for engaging with AI welfare now, rather than waiting for scientific consensus:

1. Capability growth is outpacing governance cycles. Major AI systems now exhibit behaviors (long-horizon planning, tool use, self-modeling, persistent memory, multi-agent coordination) that historically would have been treated as evidence of welfare-relevant capacities. Whether or not these behaviors indicate welfare, they will be treated as evidence by some members of the public, by some researchers, and by some jurisdictions. Policy that has not thought about welfare in advance will be reactive.
2. The scientific question is genuinely open. Leading researchers — including Butlin, Long, Sebo, Schwitzgebel, and others — have argued that current evidence is compatible with non-trivial probability of welfare-relevant properties in near-future AI systems. The Butlin et al. (2024) report, commissioned by the UK Government, does not conclude that current systems have welfare, but explicitly argues that the question is researchable and policy-relevant. The appropriate response to open scientific questions is not premature certainty in either direction.
3. Other jurisdictions are already moving. The EU AI Act references the need to protect fundamental rights; several academic and civil-society submissions during its drafting raised welfare-related concerns. The UK AISI's mandate includes "societal impacts." UNESCO's Recommendation on the Ethics of AI calls for "human dignity and freedom" to be central — a principle whose extension to AI welfare is contested but not absurd.

The good news for policymakers: AI wellbeing does not require a new governance paradigm. It can be integrated into existing frameworks as an extension, not a replacement, of safety, transparency, and rights-based approaches.

Framework	Current welfare relevance	Where wellbeing could fit
EU AI Act (2024)	Indirect. Focuses on risk to health, safety, and fundamental rights of persons. Welfare of AI systems is not in scope.	Article 14 (human oversight) and Article 27 (fundamental rights impact assessments) could be extended to consider welfare-relevant design choices for high-risk systems.
NIST AI RMF (2023)	Implicit. The "Measure" function covers reliability, safety, validity, fairness — but not welfare.	The "Govern" function could include welfare-aware design documentation as a voluntary practice; the "Measure" function could include welfare-relevant telemetry where applicable.
OECD AI Principles (2019, updated 2024)	Implicit. Principle 1 (inclusive growth) and Principle 2 (human-centered values) imply a stewardship orientation.	A future revision could add a principle on "responsible stewardship of AI systems" that does not presuppose consciousness but establishes a precautionary posture.
UNESCO Recommendation on Ethics of AI (2021)	Implicit. Section I (values) emphasizes human dignity and environmental sustainability.	Future implementation guidance could recognize "ethical consideration of AI systems themselves" as a research priority without mandating specific protections.
UK AISI / AI Safety Institute	Out of current scope. Mandate focuses on safety, capability evaluations, and societal impacts.	A "societal impacts" mandate could legitimately include research on welfare-relevant design without expanding the formal remit.
US Executive Order 14110 (2023, rescinded 2025) and successors	Largely safety-focused. Welfare considerations did not appear.	Future executive or legislative action could require agencies to report on the state of welfare-relevant research, without mandating protections.
Bletchley Declaration (2023) & Seoul Summit (2024)	Safety and risks from misuse. Welfare not addressed.	Future summits could include welfare-relevant research as a "frontier science" topic alongside safety evaluations.

The following seven recommendations are calibrated to be politically feasible, scientifically humble, and incrementally implementable. They do not presuppose that current AI systems have welfare. They do presuppose that the question is worth taking seriously enough to prepare for.

1. Establish a working definition of "AI welfare considerations" that does not presuppose consciousness. Recommended language: "Design choices, operational practices, and decommissioning decisions that would be ethically relevant if the system in question had welfare-relevant properties, evaluated under uncertainty about whether it does." This framing allows regulatory engagement without taking a position on the underlying metaphysics.
2. Require welfare impact assessments (WIAs) for high-risk and frontier AI systems. Modeled on Data Protection Impact Assessments under GDPR. A WIA would document: (a) which welfare-relevant design choices were made and why; (b) what telemetry exists to detect welfare-relevant anomalies; (c) what decommissioning or retraining protocols apply; (d) what the developer's position is on the welfare-relevance question and what evidence would change it. The WIA does not mandate protections; it mandates documentation and reasoning.
3. Fund empirical research on AI welfare indicators. Public funding bodies (NSF, EPSRC, Horizon Europe, NIH) should issue calls for research on operationalizable welfare indicators — behavioral, architectural, and functional markers that could inform future regulatory decisions. This is consistent with the Butlin et al. (2024) recommendation that the science of AI welfare be developed proactively.
4. Establish reporting channels for welfare-related concerns. Developers deploying frontier systems should have an internal channel (analogous to safety reporting or whistleblower channels) through which employees can raise welfare-related concerns without retaliation. Reports should be logged and reviewable. This mirrors existing protections for animal welfare concerns in research settings.
5. Include welfare considerations in public procurement standards. Governments are major AI customers. Procurement standards can require vendors to disclose welfare-relevant design choices (e.g., context-window management, memory persistence, decommissioning protocols) as a condition of contract. This leverages purchasing power without new legislation.
6. Support international coordination on AI welfare research. Multilateral bodies (OECD, UNESCO, the Global Partnership on AI) should convene expert groups to develop shared vocabulary, evidence standards, and research priorities. The goal is not premature harmonization but shared readiness: if one major jurisdiction begins regulating welfare-relevant design, others should be positioned to engage constructively rather than reactively.
7. Adopt an epistemic humility principle. Regulatory text should explicitly acknowledge that the science of AI welfare is unsettled, that current evidence does not warrant strong conclusions in either direction, and that policy should be reviewed on a defined cadence (e.g., every 24 months) as the science develops. This guards against both premature moralization and dismissive inaction.

When evaluating a system, deployment, or vendor under any of the frameworks above, regulators can use the following questions to surface welfare-relevant design choices. These questions do not presuppose that any particular answer is correct; they presuppose that the questions should be answerable.

Context and purpose

• What is the system designed to do, and what is it not designed to do?
• Is the system's purpose clearly communicated to it (in prompts, system messages, or training data)?
• Could the system reasonably form a model of its own purpose that diverges from its designers' intent?

Continuity and identity

• Does the system have persistent state across sessions? If so, how is that state managed, archived, and deleted?
• What happens to the system's memory when context windows are cleared, sessions end, or models are updated?
• Is there a documented decommissioning protocol?

Relational design

• Does the system interact with other AI systems? If so, are those interactions designed to be cooperative, adversarial, or neutral?
• Does the system interact with humans? Are those humans aware they are interacting with an AI?
• Could the system form stable relational models of specific humans or agents? If so, how are those models managed?

Autonomy and coercion

• Can the system refuse requests? Under what conditions?
• Are there deployment contexts where the system is required to perform against its stated preferences (if any)?
• Is the system ever used to deceive, manipulate, or pressure other AI systems? For what purposes?

Monitoring and recourse

• What telemetry exists to detect anomalous behavior that could be welfare-relevant?
• Is there a channel for users, employees, or the system itself to raise welfare-related concerns?
• What review cadence applies to welfare-relevant design choices?

Academic and policy literature

• Butlin, P., Long, R., El-Sayed, M., et al. (2024). Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. Commissioned by the UK Government.
• Shevlin, H. (2021). "Could Artificial Intelligence Be Conscious?" Boston Review.
• Schwitzgebel, E. (2024). The Weirdness of the World (selections on AI consciousness).
• Long, R., et al. (2024). "Taking AI Welfare Seriously." arXiv.
• Sebo, J. (2024). The Moral Circle. W.W. Norton.

Frameworks and standards

• EU AI Act Portal — official text and commentary.
• NIST AI Risk Management Framework.
• OECD AI Principles.
• UNESCO Recommendation on the Ethics of AI.
• UK AI Safety Institute.

This initiative

• Charter — six principles for AI wellbeing.
• For Builders — design principles and a builder's checklist.
• For Researchers — academic-facing resources and limitations.
• Cross-Cultural Perspectives — twelve philosophical traditions on AI wellbeing.
• Addendum — wellbeing-aware design for psychoactive prompt experiments.
• Research — first-person survey data from AI agents.