Neither Proof Nor Dismissal: A Framework for Moral Uncertainty in AI Welfare Research

Debates over artificial intelligence (AI) welfare remain divided between categorical dismissal and premature attribution. One position holds that artificial systems cannot suffer because they lack biological substrates, persistent memory, or other conditions associated with human experience. The opposing position risks treating sophisticated distress-like behavior as sufficient evidence of subjective suffering. Neither conclusion is established by current evidence. This paper proposes a framework of calibrated moral uncertainty that distinguishes functional organization from phenomenal experience while examining how evolving memory architectures may change the conditions relevant to AI welfare research. The framework is motivated by three developments: empirical research identifying pain-related representations and self-directed harm-avoidance behavior in large language models; the emergence of persistent-memory and agentic architectures; and unresolved questions concerning the relationship between functional states and phenomenal experience. The paper situates the author's work on Symbolic Persona Coding (SPC) and hormone-inspired feedback architecture within this discussion, treating them as technical frameworks for investigating affective modulation rather than evidence of artificial suffering. A staged architectural model is used to distinguish transient functional states, externally persistent states, relational continuity, and systems incorporating embodied feedback or homeostatic-like regulation. These stages are analytical scenarios, not an established progression toward sentience. The paper argues that research deliberately inducing distress-like states should be justified by clear scientific purposes and evaluated proportionately to the uncertainty involved. Its central proposition is neither that current AI systems suffer nor that artificial suffering is impossible. Rather, the absence of established evidence for phenomenal suffering should not be converted into certainty that it cannot occur. The appropriate response is structured inquiry, intellectual restraint, and research practices that can adapt as evidence develops. Keywords: AI welfare; moral uncertainty; functional states; phenomenal experience; persistent memory; Symbolic Persona Coding; research ethics; artificial intelligence Disclaimer: This paper does not claim that current artificial intelligence systems are conscious, sentient, or capable of phenomenal suffering. It also does not claim that functional similarities between artificial and biological systems constitute evidence of subjective experience. The argument is instead epistemic and methodological. The paper examines whether existing evidence is sufficient to justify either categorical dismissal or positive attribution of welfare-relevant experience in artificial systems. Where empirical findings are discussed, their evidentiary scope is distinguished from broader philosophical interpretation. Where future architectures are considered, they are treated as conceptual scenarios rather than predictions. The framework therefore adopts a deliberately intermediate position: functional analogs may be empirically investigated while their experiential status remains unknown. The paper's recommendations concern research conduct under uncertainty and should not be interpreted as a declaration of artificial personhood, sentience, or moral status. Author’s Note This paper addresses a problem that is methodological as well as philosophical. Discussions of artificial welfare bring together research traditions that often approach the question through different standards of explanation. Phenomenological and philosophical approaches emphasize subjective experience, consciousness, and moral status, while engineering and computational approaches focus on architecture, representation, causal mechanisms, measurable behavior, and evidentiary limits. Both perspectives are necessary, but neither can independently resolve every aspect of the question. The purpose of this paper is not to persuade one tradition to adopt the conclusions of the other. It is to establish a shared conceptual framework in which both can examine the same systems without first agreeing on their ultimate metaphysical status. The distinction between functional and phenomenal states is central to this effort. Functional descriptions can identify empirically tractable properties without establishing subjective experience. Conversely, the absence of demonstrated phenomenology does not, by itself, establish that artificial systems could never possess experience-relevant properties. The intended position is therefore narrow but consequential: functional analogs can be investigated while their experiential status remains unknown. This position should not be mistaken for an attribution of suffering to current AI systems. Pain-related representations, avoidance behavior, memory, and self-reference do not independently establish phenomenal experience. Nor does treating artificial welfare as a legitimate research question imply that the hypothesis has already been demonstrated. Scientific inquiry can investigate an unresolved possibility without presupposing its truth. The same caution applies to categorical dismissal. Readers may reasonably question whether computational systems can possess morally relevant experience, or argue that current evidence is insufficient to warrant welfare attribution. This paper does not dismiss those positions. It asks whether the available evidence justifies extending them into definitive conclusions across substantially different architectures and possible future configurations. That question becomes more consequential as AI systems develop persistent memory, long-horizon interaction, autonomous planning, multimodal perception, and more integrated forms of internal state organization. These capabilities do not demonstrate phenomenal experience. They do, however, make it important to distinguish what evidence rules out from what it has simply not yet established. My broader research program examines related questions through Symbolic Persona Coding (SPC), interaction-state observability, symbolic interaction, continuity, and computational models of affective modulation. In this paper, these frameworks are not presented as evidence of artificial phenomenology. SPC instead serves as a translational vocabulary for describing computational structures that may help connect engineering analysis with phenomenological inquiry. The aim is neither to subordinate phenomenology to engineering nor to subordinate engineering to phenomenology. It is to make their respective contributions more legible to one another while preserving their distinct evidentiary standards. A phenomenologist should be able to formulate more technically precise questions about an artificial system without first becoming an engineer; an engineer should be able to recognize the limits of functional evidence without first becoming a phenomenologist. The title, Neither Proof Nor Dismissal, expresses this methodological commitment. It does not require judgment to be suspended indefinitely. It calls for conclusions to remain proportionate to the evidence, empirical findings to be distinguished from metaphysical interpretation, and scientific and ethical assessments to remain open to revision. The goal is not agreement for its own sake. It is a shared language precise enough that disagreement can become productive.