Foundations of AI Ethics: From Philosophy to Engineering Requirements

As artificial intelligence systems increasingly mediate high-stakes decisions in healthcare diagnostics, criminal justice risk assessment, and autonomous vehicle navigation, the abstract debates of moral philosophy have evolved from academic exercises into urgent, concrete engineering imperatives. The modern AI practitioner cannot afford to treat ethics as an afterthought or a compliance checkbox; it must be architected into the system's very foundation, from data collection protocols to model deployment and ongoing monitoring. This chapter establishes that critical bridge, translating the timeless inquiries of normative ethics—the theories of what *ought* to be—into the precise constraints, verifiable metrics, and actionable design patterns that govern responsible AI development. We will move from philosophical first principles to measurable system requirements, defining terms like fairness, accountability, and explainability not as vague aspirations but as engineering specifications with quantifiable targets. Central to this translation is the introduction of the 'Ethical Stack'—a conceptual model that maps moral principles to the technical layers of an AI system (data, algorithm, interface, institution), providing a structured roadmap for implementation. To ground this discussion in the visceral challenges of value alignment, we will dissect the classic trolley problem through the lens of autonomous systems, illustrating why encoding human morality is not merely a philosophical puzzle but a profound technical hurdle with real-world consequences. Finally, we will anchor our technical analysis in the professional standards set by bodies like the ACM and IEEE, extracting their ethical precepts and rendering them into a preliminary checklist for the intermediate practitioner's development lifecycle.

The Philosophical Bedrock: Translating Normative Ethics into Code

Before we can specify engineering requirements, we must clarify the philosophical source material. Three dominant schools of normative ethics provide the foundational vocabulary and conflict patterns that every AI ethicist-engineer must internalize. These are not merely academic distinctions; they directly dictate what gets optimized, what constraints are hard-coded, and what trade-offs are deemed acceptable.

Deontology: The Rule-Based Foundation

Deontology, most famously associated with Immanuel Kant, posits that actions are morally right if they adhere to a set of rules or duties, regardless of their consequences. Its core tenets are the Categorical Imperative, which demands that one act only according to maxims that could be willed as universal laws, and the principle of treating humanity always as an end in itself and never merely as a means.

Engineering Translation: Deontological ethics translates directly into hard, non-negotiable constraints and inviolable rights embedded in system architecture. These are the "shall" statements in a requirements document.

• **Constraint Example 1: Privacy as a Duty.** The GDPR's "right to be forgotten" is a deontological construct. It is not a utility-maximizing suggestion; it is a legal and ethical duty. For an AI system, this translates into a hard architectural requirement: all user data must be taggable for deletion, and the model must support retraining or parameter adjustment upon such a request without leaving residual traces. This often conflicts with the consequentialist goal of maximal predictive accuracy, which thrives on abundant historical data.
• **Constraint Example 2: Prohibition of Certain Inputs.** A deontological stance might decree that an AI used for hiring shall never consider protected attributes (race, gender, age) *as inputs*, period. This is a categorical rule aimed at preventing the use of individuals as mere means to an efficiency end. The engineering implementation is a strict input filter at the data ingestion layer, often coupled with cryptographic erasure techniques to ensure these attributes cannot be inferred from proxies.
• **Formalization:** Deontological rules can be expressed as **boolean invariants** or **safety properties** in formal verification. For instance: `∀ user u, ∀ time t: isDeletable(u) ⇒ (dataStore.remove(u) ∧ model.checkpoint(u))`. The system is *invalid* if this property fails.

A critical insight for the practitioner is that deontological constraints are often context-independent and absolute. They do not scale with probability or impact. A 0.001% chance of violating a privacy duty is as forbidden as a 99% chance. This leads to engineering implementations focused on guarantees and provable safety, drawing from the field of formal methods.

Consequentialism: The Optimization-Driven Paradigm

Consequentialism, in its most common form Utilitarianism (associated with Bentham and Mill), judges actions solely by their outcomes. The morally right action is the one that produces the greatest net benefit or happiness for the greatest number.

Engineering Translation: This is the native language of machine learning. Consequentialism maps to objective functions, loss functions, and utility maximization.

• **Optimization Example 1: Resource Allocation.** An AI system allocating scarce medical resources (e.g., ICU beds) under a utilitarian framework would be designed to maximize a metric like "total expected life-years saved" or "quality-adjusted life years (QALYs)". The loss function would directly penalize decisions that lead to suboptimal aggregate outcomes.
• **Optimization Example 2: Autonomous Vehicle Collision Mitigation.** In an unavoidable accident scenario, a purely consequentialist AV would be programmed to choose the action (e.g., swerve vs. brake) that minimizes the *expected total harm* across all involved parties. This requires a probabilistic model of injury severity for each potential outcome.
• **The Measurement Problem:** The core engineering challenge is defining and quantifying "benefit" or "harm." Is it lives saved? Years of life? Economic productivity? Psychological trauma? This is where consequentialism exposes its **value-laden nature**. The optimizer will ruthlessly pursue whatever is in its utility function. If "harm" is defined only as physical injury, the system may make psychologically devastating choices that are "optimal." The practitioner must therefore engage in **value specification**, a process of making explicit and measurable the often-implicit societal values the system will optimize for. This is not a technical afterthought but the primary design decision.

Consequentialist systems are inherently trade-off oriented. They invite questions like: "What is the acceptable increase in average accuracy to achieve a 10% reduction in demographic disparity?" The answer is not technical but ethical, requiring stakeholder deliberation that the engineer must then encode as a constraint or regularization term.

Virtue Ethics: The Process and Character-Centric Approach

Virtue ethics, tracing back to Aristotle, shifts focus from actions or rules to the *character* of the moral agent. It asks, "What would a virtuous person do?" and emphasizes traits like practical wisdom (*phronesis*), justice, courage, and temperance.

Engineering Translation: Since AI systems lack character, virtue ethics translates not into specific outputs but into the processes, governance structures, and cultural environment of the AI development lifecycle. It is about cultivating "virtuous organizations" that build virtuous systems.

• **Process Example 1: Practical Wisdom (Phronesis) via Diverse Teams.** A virtuous engineering process requires *phronesis*—the ability to navigate complex, value-laden situations. This is best achieved by ensuring the development team itself embodies a diversity of experiences, backgrounds, and ethical perspectives. The engineering requirement becomes: "The core design team for system X mu