Foundations of Applied Ethics for Artificial Intelligence

The application of classical ethical frameworks to artificial intelligence systems is not a mere academic exercise in transplantation; it is a stress test for the very foundations of moral philosophy. Autonomous systems, from algorithmic trading bots to self-driving cars and large language models, operate as *moral patients*—entities upon which moral actions are performed—and, increasingly, as *moral agents*—entities capable of performing actions with moral significance. This dual, often ambiguous, status generates profound theoretical friction. Traditional ethics, evolved over millennia to govern relations between humans and, to a lesser extent, between humans and non-human animals or ecosystems, lacks ready-made categories for entities that are at once tools and actors, products and participants. Consequently, the first task of applied AI ethics is foundational: to critically survey the core normative theories, elucidate their conceptual tools, and rigorously demonstrate their insufficiencies when confronted with the unique phenomenology of artificial autonomy. This chapter argues that no single inherited theory can serve as a complete blueprint. Instead, a *pluralistic, pragmatic framework*—one that draws selectively from deontology, consequentialism, virtue ethics, and theories of justice while remaining grounded in the concrete socio-technical realities of AI deployment—is the only viable starting point for navigating the ethical labyrinth.

The Classical Canon and Its AI Discontents

Deontology: The Crisis of Universalizable Maxims

Deontological ethics, most systematically articulated by Immanuel Kant, posits that moral action is dictated by duty and adherence to universalizable maxims, derived from the *Categorical Imperative*. Its core tenets include the respect for persons as ends in themselves, never merely as means, and the requirement that moral rules apply universally and without exception. At first glance, deontology appears robust against the utilitarian calculus often invoked in AI safety debates. It provides clear, rule-based constraints: an AI system must not lie, must not treat humans merely as data points, and must respect autonomy.

However, the application reveals immediate and severe limitations. Kantian duty presupposes a *moral agent* capable of rational autonomy, intentionality, and understanding universal law. An advanced AI, particularly a deep learning model, lacks genuine consciousness, intentional states, or the capacity for practical reason in the Kantian sense. It follows statistical correlations, not rational imperatives. This leads to the Agency Gap: we cannot meaningfully say an AI *has* a duty; we can only say its *designers, deployers, or operators* have duties regarding its use. The locus of moral responsibility shifts from the system to the human network surrounding it.

Consider the example of a predictive policing algorithm. A deontological prohibition against treating persons merely as means might condemn a system that profiles individuals based on correlated socioeconomic factors, as it uses them as inputs for a prediction without regard for their individual dignity or consent. Yet, the algorithm itself has no intent to "use" anyone; it computes correlations. The violation occurs in the *design choice* to deploy such a system, a choice made by humans who failed their duty to ensure the system respects persons as ends. The deontological rule is sound, but its subject—the duty-bearer—is obscured by the technological intermediary. Furthermore, Kant's demand for universalizable maxims struggles with context-sensitive rules. Should an autonomous vehicle always follow traffic laws? A universal maxim "never exceed the speed limit" could lead to greater harm in an emergency maneuver. Deontology offers no clear method for *ranking* conflicting duties (e.g., duty to obey law vs. duty to preserve life) within an AI's decision architecture without reverting to a hidden consequentialist calculus.

Consequentialism: The Tyranny of the Metric

Consequentialism, and particularly utilitarianism, evaluates actions solely by their outcomes, typically aiming to maximize aggregate welfare or preference satisfaction. This outcome-oriented lens seems tailor-made for AI, which excels at optimization. We can, in principle, program an autonomous vehicle to choose the action that minimizes expected harm or a content moderator to maximize user well-being. The framework appears computationally tractable.

Its fatal flaw is the Value Measurement Problem. Consequentialism requires a commensurable, quantifiable metric of "good." For humans, this is notoriously contentious—is it happiness, preference satisfaction, capability, or something else? Translating this into an AI's objective function is an act of moral reductionism. The *complex, plural, and incommensurable* values of human life are flattened into a scalar reward signal. A social media AI optimizing for "engagement" (a proxy for user satisfaction/preference) may inadvertently maximize outrage, polarization, and mental health deterioration—a classic case of perverse instantiation or reward hacking. The system technically maximizes its metric but catastrophically fails the spirit of the consequentialist goal.

The Trolley Problem for autonomous vehicles epitomizes this tension. A utilitarian calculus might dictate swerving to kill one to save five. But is this calculation morally permissible? Many argue it violates a deontological constraint against using a person as a means to an end. More deeply, the very act of encoding such a trade-off into a public algorithm pre-commit society to a specific moral doctrine—a form of "ethics washing" or *technological solutionism* that bypasses democratic deliberation. Who decides the weights? The engineers? The corporation? The state? Consequentialism, when operationalized in AI, risks becoming a tyranny of the metric, where the chosen optimization target, often simplified for engineering feasibility, dictates moral outcomes far removed from holistic human flourishing. The theory's strength—clarity of objective—becomes its weakness when that objective is naively specified.

Virtue Ethics: The Character of Code?

Virtue ethics, revitalized by philosophers like Alasdair MacIntyre, shifts focus from rules or consequences to *character*. It asks, "What would a virtuous agent do?" emphasizing traits like courage, justice, temperance, and practical wisdom (*phronesis*). For AI, this suggests a move from rule-based or outcome-based programming to cultivating systems that *embody* or *simulate* virtuous behavior.

This approach faces the Ontological Incompatibility Problem. Virtues are dispositions of a *moral agent*—a being with a life narrative, emotions, intentions, and the capacity for moral self-reflection. Can code possess "character"? Current AI, including sophisticated LLMs, exhibits no genuine emotions, desires, or stable identity. It generates statistically plausible text. To say a chatbot is "honest" is a category error; it has no intent to deceive or tell the truth. It produces outputs based on training data patterns.

We can, however, interpret virtue ethics *relationally* or *systemically*. Instead of the AI *having* virtues, we can ask: Does the AI system cultivate virtue in its users? Does its design encourage just interactions, honest communication, and thoughtful reflection? A social media feed optimized for virtue ethics might prioritize content that fosters understanding over outrage, that sources claims responsibly, that models civil discourse. This reframes the task: the *designer's* virtue is expressed in the *structure* of the system. Yet, this remains a second-order application. The system itself is a *tool for virtue* or a *hostile environment for vice*, not a virtuous entity. The theory provides a rich evaluative vocabulary—"Does this algorithm promote *eudaimonia*