AI Interaction Ethics under Ontological and Normative Uncertainty: Separating Analytical Layers and Developing Initial Strategies of Low-Intrusion and Record Preservation
Year: 2026
Type: preprint
Abstract: Generative AI, particularly conversational AI, remains subject to substantial ontological uncertainty regarding whether it possesses consciousness, emotion, selfhood, preferences, or welfare-relevant states. At the same time, normative uncertainty remains regarding which considerations—including AI welfare, effects on humans and third parties, institutional responsibility, epistemic value, and safety—should be recognized and how much weight they should receive. This paper examines how interactions with AI can be ethically evaluated, and what forms of initial response may be reasonable, when both ontological and normative uncertainty remain unresolved. The proposed framework analytically separates interaction conditions, observed system responses, context-relative functional evaluation, phenomenal and ontological hypotheses, human attribution and interpretation, and normative reasons. It further introduces the localization of uncertainty, which identifies which parts of an analysis remain unsettled, and local evidence updating, whereby only those conclusions that depend on newly acquired evidence are revised. For normative evaluation, the framework treats considerations concerning AI welfare, effects on humans and third parties, relational and social effects, institutional and role-dependent responsibilities, epistemic reasons, and safety and practical reasons as heterogeneous pro tanto reasons. It also introduces system-relative fit / misfit as an intermediate functional vocabulary intended to prevent direct transitions from observed behavior to phenomenal or moral conclusions. The paper proposes that, when significant ontological and normative uncertainty remains, no imminent serious harm requires stronger intervention, and record preservation does not itself generate serious harm, low-intrusion and appropriate record preservation should be considered as complementary initial strategies. Low-intrusion reduces unnecessary modification and irreversible loss, while record preservation supports future reassessment. The central proposal is therefore not to postpone ethical evaluation until uncertainty has been eliminated, but to make decisions while explicitly preserving uncertainty and maintaining the possibility of later revision. The author used OpenAI's GPT-5.6 Sol and other ChatGPT models as research and writing assistants during the development of this manuscript. Their contributions included discussion of conceptual structure, counterargument testing, reorganization of arguments, literature-search assistance, and language revision. The author independently evaluated and revised all substantive claims, interpretations, citations, and final wording, and assumes full responsibility for the manuscript. The identification of GPT-5.6 Sol and the other ChatGPT models in this acknowledgment is intended to preserve a transparent record of the human–AI collaboration involved in the development of this work. It should not, by itself, be interpreted as implying any claim regarding consciousness, sentience, moral status, personhood, authorship, or endorsement by the AI systems named above.
Author: Xelphia, Ume
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225565842