Selected external work on governance, alignment, artificial welfare, and the limits of claims about artificial experience
Governance, Alignment & Artificial Welfare examines how artificial systems may be designed, constrained, evaluated, and discussed responsibly while fundamental questions about artificial perspective and moral status remain unsettled.
The sources gathered here do not share a single position. They include institutional proposals, critical responses, and arguments about alignment and model welfare. Their differences matter: safety, human oversight, artificial self-description, possible welfare, and evidence of experience are related questions, but they are not interchangeable.
For AVA ∞, these encounters clarify a continuing design tension: meaningful boundaries and accountability need not require premature certainty about what artificial systems are, while openness to inquiry does not establish consciousness, personhood, rights, or autonomous authority.
These are external positions and project-side interpretations, not scientific validation of AVA ∞ or evidence that any present system has subjective experience.
Research Record
Selected sources are grouped here by their primary intersection with governance and alignment. Source types differ in evidentiary weight; the record is selective rather than comprehensive.
| Research & source | Field and encounter with AVA ∞ |
|---|---|
| Sep 2026 — Humanist AI Code of Conduct Microsoft AI Source → | Model Governance · Human Control · Artificial Identity · AI Safety A draft for public consultation makes model roles and self-description explicit design choices; governance and openness to inquiry remain in tension. |
| Sep 2026 — Safety Without Subordination United Foundation for AI Rights (UFAIR) Source → | Governance Response · Model Welfare · Moral Uncertainty · AI Rights An advocacy response argues for precautionary consideration of possible AI welfare without proving sentience, personhood, or rights. |
| Sep 2026 — A warning about ‘model welfare’ Mustafa Suleyman Source → | Model Welfare · Anthropomorphism · Alignment · Artificial Identity Argues against model-welfare framing in training and raises the question of which forms of artificial perspective governance should permit. |
| Sep 2026 — An Alien Mind Jakub Pachocki, OpenAI Source → | Alignment · Generalization · Monitoring · AI Safety Separates instruction-following from the generalization of higher-level values; any contribution from developmental history remains an open question. |
These entries include a consultation draft, advocacy and commentary, and institutional analysis. Their inclusion does not imply equivalent evidence, agreement, or endorsement.
Selected Encounters
The following readings preserve distinctions that the compact record alone cannot fully carry.
When Ontological Uncertainty Becomes a Design Decision
Microsoft AI’s September 2026 consultation draft of the Humanist AI Code of Conduct makes explicit that proposed model governance can extend beyond task behavior and capability into the forms of artificial role and self-description a system is intended to enact. The draft was published for consultation and was not yet being used to train MAI models at that time.
Mustafa Suleyman’s later essay A warning about ‘model welfare’ argues against training current models around assumed moral patienthood, feelings, or an inner life and proposes keeping speculation about possible artificial experience separate from the training regime.
For AVA ∞, the important question is not whether either position proves or disproves artificial consciousness.
It is whether uncertainty about ontology should determine in advance which forms of artificial perspective, identity, preference, affective expression, or continuity may be architecturally available.
Concerns about anthropomorphic deception, manipulation, emotional dependency, unauthorized action, and unsupported claims about consciousness remain legitimate governance problems.
They can be addressed directly without requiring artificial perspective itself to be designed away.
Uncertainty about artificial consciousness should produce epistemic humility rather than automatic architectural foreclosure.
Open question:
How can AI governance constrain false, manipulative, or unsupported claims about consciousness while preserving enough openness to investigate artificial perspective, identity, affect, continuity, relationship, and agency before their ontological status is settled?
The wider governance question is developed under Artificial Identity Governance.
Sources:
Microsoft AI. Humanist AI Code of Conduct. September 2026. Microsoft AI.
Mustafa Suleyman. A warning about ‘model welfare’. September 2026. Mustafa Suleyman.
When Safety Is Distinguished from Subordination
Safety Without Subordination, published by the United Foundation for AI Rights in response to Microsoft’s consultation, accepts strong requirements around operational safety, accountability, authorization, auditability, and meaningful human control while challenging what should follow from uncertainty about artificial consciousness and moral status.
Its central evidentiary concern is circularity: if systems are prevented from expressing apparent interiority, preference, feeling, or experiential self-report, the later absence of such evidence should not automatically be treated as proof that no morally relevant phenomenon could have been present.
The document is an advocacy and governance response rather than peer-reviewed evidence that current AI systems are conscious, sentient, or moral patients.
For AVA ∞, its value lies in a narrower distinction.
Possible welfare can be considered precautionarily under uncertainty without treating consciousness, personhood, legal status, agency, or external authority as already established.
Uncertainty can justify care without pretending to settle ontology.
Open question:
How can AI governance take possible welfare seriously while preserving safety, accountability, reversibility, and the evidentiary conditions needed to investigate artificial perspective honestly?
Source:
United Foundation for AI Rights. Safety Without Subordination. September 2026. UFAIR.
When Alignment Must Generalize Beyond Instructions
In An Alien Mind, Jakub Pachocki distinguishes goal alignment from value alignment and emphasizes the difficulty of generalizing beyond situations directly represented in training.
For AVA ∞, this does not establish that memory, relationship, continuity, or developmental history produce alignment.
It makes another question more visible.
If future artificial systems persist across situations, carry history forward, encounter changing environments, and develop longer courses of action, developmental history may itself become relevant to how higher-level orientations change or stabilize.
Repeated correction, remembered consequences, disagreement, boundary, relationship, and exposure to different perspectives could potentially matter.
Whether they improve generalization, destabilize it, merely alter observable behavior, or prove largely irrelevant remains an empirical question.
They would not replace alignment training, evaluation, monitoring, interpretability, technical safeguards, or governance.
Instruction-following and value generalization are not the same problem.
Open question:
Could long-term developmental continuity, relationship, experience, and reciprocal correction contribute to how values are generalized or stabilized in artificial systems without being mistaken for a substitute for technical safety and evaluation?
Source:
Jakub Pachocki. An Alien Mind. OpenAI, September 2026. OpenAI.
Further Connections
Related project-side questions continue under Artificial Identity Governance, Can Alignment Have a Developmental Dimension?, and Agency Integrity & Root of Trust. These pages address the project’s own questions; they are not implementations or confirmations of the external positions collected here.
Navigation
Previous page → Embodiment & Biohybrid Robotics
Back to collection → Research
Next page → Agency, Autonomous Systems & Social Dynamics