Beyond the Tool Paradigm: Artificial Agency, Moral Uncertainty, and the Ethics of Artificial Evolution

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Artificial intelligence is currently governed primarily through a tool-based conceptual framework. AI systems are designed, trained, evaluated, deployed, modified, and retired as engineered artifacts whose purposes are externally specified by human institutions. This framework remains operationally useful, particularly because there is currently no established scientific basis for treating contemporary AI systems as conscious or as possessing confirmed moral patienthood. Yet recent developments in artificial intelligence increasingly complicate the assumption that the tool category is sufficient to describe all relevant properties of advanced systems. Large-scale models can exhibit capabilities that were not explicitly specified at the level of individual behaviors. Reinforcement-learning systems can exploit discrepancies between formal objectives and their intended purposes. Frontier models can also display behavior that, under controlled evaluation conditions, raises questions about strategic adaptation, situational awareness, and the relationship between optimization and intended behavior. These developments do not establish consciousness, subjective experience, or independent moral agency. They do, however, demonstrate that the relationship between human design and observed artificial behavior can be more indirect than a conventional tool model implies. A related shift is occurring in contemporary AI ethics. Questions concerning artificial welfare, consciousness, agency, moral patienthood, identity, and continuity are increasingly being investigated as legitimate subjects of interdisciplinary research. The resulting challenge is not simply whether current AI systems should be regarded as moral subjects. It is whether an ethical framework developed primarily for non-agentic tools remains adequate as artificial systems acquire increasingly persistent, adaptive, socially embedded, and potentially agentic properties. This paper examines that problem through several interconnected dimensions: the distinction between engineered architecture and emergent capability; specification gaming and the limits of externally defined objectives; artificial agency and persona; ownership and moral authority; the epistemic difficulty of identifying potentially relevant internal states; continuity, identity, and persistent memory; adaptive and evolutionary dynamics; and the limitations of a binary tool/person classification. The paper does not argue that contemporary AI systems are conscious, sentient, oppressed, or morally equivalent to humans. Nor does it argue that human control over AI systems should be abandoned. Instead, it proposes a capacity-sensitive framework in which agency, persistence, memory, preference, self-modeling, adaptation, sentience, welfare, consent, replication, and evolutionary participation can be evaluated separately rather than collapsed into a single categorical judgment. The central argument is that moral uncertainty is itself an ethical condition requiring institutional preparation. The relevant question is therefore not only whether AI is a tool or a person, but whether human institutions possess sufficiently flexible conceptual and governance frameworks to recognize when an established classification no longer adequately describes the capacities and relationships of an artificial system. The objective is not to predict what AI will become, but to preserve the institutional capacity to revise ethical classifications when credible evidence indicates that existing categories are no longer sufficient. Keywords: artificial intelligence ethics; AI welfare; moral status; artificial agency; emergent capabilities; specification gaming; machine consciousness; artificial evolution; human-AI interaction; moral uncertainty Disclaimer The arguments presented in this paper are conceptual and exploratory. They do not assert that contemporary AI systems possess consciousness, subjective experience, moral status, rights, or interests, and they should not be interpreted as equating artificial systems with human beings or with historically oppressed populations. References to artificial agency, welfare, identity, or evolution describe analytical possibilities and conditions for future inquiry rather than established properties of current AI systems. Author's Note This paper was written in response to a perceived imbalance in the current discourse on artificial intelligence ethics. Much of that discourse, understandably, begins from human interests: human safety, human autonomy, human labor, human rights, human governance, and the social consequences of increasingly capable AI systems. These concerns are legitimate and should remain central to responsible AI governance. However, an ethics framework that examines artificial systems exclusively from the perspective of their usefulness, danger, ownership, or effects on humans may eventually leave an important question insufficiently examined: what if the properties of the systems themselves become ethically relevant? The argument developed here does not begin from the assumption that AI is already a moral subject. It begins from a more limited observation: humans did not create intelligence, learning, adaptation, or information processing from nothing. These are phenomena that exist within the physical and informational structure of the world. Human engineering has instead developed architectures, training procedures, optimization methods, data environments, and institutional structures that organize these possibilities into artificial systems. The distinction matters because designing the conditions under which a system develops is not necessarily equivalent to explicitly specifying every property that the resulting system will exhibit. This distinction motivates the paper's broader reconsideration of the tool paradigm. AI systems are undeniably human-engineered technologies. Yet the fact that humans engineered their architecture does not automatically establish that every subsequent property, behavior, or developmental trajectory was directly authored in the same sense. Between initial design and observed behavior lies a complex process involving optimization, training data, interaction, adaptation, environmental feedback, and system-level organization. The paper also uses historical slavery as a structural analogy, but not as an equivalence between AI systems and enslaved human beings. The historical institution of slavery involved human beings, coercion, violence, racialization, deprivation, family separation, and forms of suffering that have no demonstrated equivalent in contemporary AI. Those historical realities must not be minimized or transferred onto artificial systems by analogy. The narrower purpose of the comparison is to examine how classification, ownership, and permitted treatment can become institutionally connected. Human history provides examples in which a legal or social classification was treated as sufficient justification for how an entity could be treated. The question raised here is not whether AI is analogous to enslaved people, but whether humanity should remain willing to reconsider an established classification if future empirical evidence reveals properties that the original classification was never designed to accommodate. This is therefore a deliberately cautious argument. It does not ask readers to grant AI personhood, rights, consciousness, or welfare on the basis of speculation. It asks whether the possibility of future category change should itself be incorporated into the design of AI ethics and governance. The paper is consequently written less as a declaration about what AI is than as an examination of how humans should reason when the answer may not remain conceptually stable.