Whose morality? Which rationality? Challenging artificial intelligence as a remedy for the lack of moral enhancement

The paper addresses whether algorithmic decision-making can achieve moral autonomy comparable to human moral agency. Within machine ethics, the author distinguishes between Moor’s implicit ethical agents (machines programmed to behave ethically) and explicit ethical agents (machines able to compute the best action using ethical principles), mapping these to artificial moral agents (AMA) and artificial autonomous moral agents (AAMA). Revisiting MacIntyre’s questions as “Whose morality?” and “Which rationality?”, the study examines how moral reasoning relates to computation and whether an ethical intelligent agent can be created. The work situates strong AI vs weak AI distinctions within moral terms (strong “moral” AI: explicit ethical agents; weak “moral” AI: implicit ethical agents) and motivates a critical analysis of three normative frameworks—Kantian, utilitarian, and virtue ethics—to probe the feasibility and limits of moral machines. The importance lies in clarifying risks of conflating algorithmic decision processes with human moral reasoning, especially under bias and the Moral Lag Problem, and in assessing prospects for moral self-update in machines.

Conceptual and comparative philosophical analysis. The author: - Frames the Moral Lag Problem and examines implications for machine ethics. - Analyzes how computation-based decision-making risks simulating ethical deliberation and how designer biases and algorithmic power shape outcomes, potentially intermingling computational errors with moral mistakes. - Tests three first-order normative approaches as models for moral machines: 1) Deontological (Powers’ Kantian machine): evaluates nonmonotonic logic, default rules, permissible maxims, and the issue of semidecidability; explores the machine’s grasp of obligation and the role/lack of moral feelings. 2) Act utilitarianism (Anderson & Anderson): investigates computing utility, uncertainty, endless calculation vs estimation, lack of clear input maxims, and challenges of measuring utility and interpersonal comparisons; examines the role of moral feelings (e.g., happiness) and their computability. 3) Virtue ethics (Howard & Muntean’s AAMA): evaluates an agent-centered, active-learning approach using soft computation (NN + EC), pattern recognition, dispositional traits (robo-virtues), black-box strategies, and bias/distributed responsibility; assesses risks of moral relativism and limitations in reading moral behavior from data. - Explores the epistemic-moral predictability gap and the feasibility of moral self-update under each framework. No empirical datasets are used; arguments proceed through logical unpacking of theories, critical synthesis, and thought-experimental evaluation of design implications.

- Computation vs estimation: Across Kantian, utilitarian, and virtue-ethical models, computation does not straightforwardly translate into moral estimation; algorithmic success does not ensure moral agency. - Risk of simulacrum: Explicit ethical agents may only simulate ethical deliberation, amplifying imperfect human morality (via bias) into an immoral hyper-reality. - Bias and error conflation: Algorithm design reflects designers’ values; institutional and socio-political contexts can reinforce norms. There is a persistent risk of conflating computational errors with moral mistakes. - Kantian model: Nonmonotonic logic is preferable to monotonic for deontological rules but suffers from failure of semidecidability. Without an ethical analogue of the third semidecidability aspect (non-halting search for membership), moral self-update is impaired. Permissible maxims as input can contradict facts without inconsistency, yet safeguarding against immoral outcomes remains problematic. Kantian machines also lack moral feelings and may not distinguish “I ought to do z” from “z ought to be the case.” - Utilitarian model: Computing utility faces unresolved problems (what is utility, how to measure it, interpersonal comparability). Algorithms risk an endless stream of calculations and, more subtly, endless estimation that reduces morality to cost-benefit analysis. Lack of clear input maxims undermines moral self-update; equating goodness with happiness is questionable, and aligning majority good with individual good is unresolved. - Virtue-ethical AAMA: The agent-centered, soft-computing approach acknowledges nonlinearity, noise, and errors, potentially aiding data correction. Yet black-boxing mental/psychological mechanisms limits moral understanding; distributed responsibility increases bias influence; robo-virtues can vary across scenarios, risking relativism or immoralism. Pattern reading may miss moral motivation and feelings, challenging robust moral self-update. - Strong vs weak “moral” AI: Both strong (explicit) and weak (implicit) scenarios are problematic. Human cognition does not equal human moral reasoning; absent moral feelings and complex motivation analogues, analogical replication fails. Moral self-update lacks guarantees due to the gap between epistemic and moral predictability. - Only humans qualify as full ethical agents in Moor’s sense; thus, achieving lasting machine moral perfection is currently unsupported.

The analysis shows that embedding first-order normative theories into machines does not resolve the core challenge: moral reasoning involves context-sensitive judgment, motivation, and often moral feelings, which are not reducible to computation. The Moral Lag Problem and designer/institutional biases further risk transforming algorithmic decision systems into simulacra of ethical deliberation rather than genuine moral agents. By examining Kantian, utilitarian, and virtue-ethical approaches, the study demonstrates shared limits: logical constraints (e.g., semidecidability) and practical/normative uncertainties (utility metrics, black-box learning) undermine reliable moral self-update. This directly addresses the research questions—“Whose morality? Which rationality?”—by revealing that no single codified rationality or imported morality suffices for machine moral autonomy. The findings imply that both strong and weak “moral” AI scenarios fail to align epistemic predictability with moral predictability, challenging the prospect of machines achieving or surpassing human-level moral agency. Responsibility for machine behavior remains human-centered, and claims of long-run moral robustness are not morally justified without mechanisms ensuring moral, not just epistemic, improvement.

The paper argues that building well-functioning AI systems for moral purposes does not make them moral machines by default. Across deontological, utilitarian, and virtue-ethical instantiations, attempts to formalize moral reasoning encounter logical and ethical hurdles (e.g., semidecidability limitations, utility measurement, and black-box virtue learning). Biases and the conflation of computational errors with moral mistakes threaten to turn moral machines into simulacra of ethical deliberation. Because human cognition differs from human moral reasoning, and due to the lack of analogues for moral feelings and complex motivation, both strong and weak “moral” AI scenarios remain doubtful. Moral self-update cannot be guaranteed as progressive or morally reliable given the gap between epistemic and moral predictability. The study concludes that only humans currently meet the criteria for full ethical agency; thus, assurances of enduring machine moral perfection are unwarranted. Future work might explore cautiously delimited, transparent, and accountable systems that assist human moral agents (e.g., Socratic assistants), investigate ethically informed nonmonotonic frameworks with safeguards, and develop governance structures addressing bias, responsibility, and override/disable mechanisms in high-stakes domains.

- Conceptual scope: The work is a philosophical/conceptual analysis without empirical validation or datasets; conclusions rely on theoretical critique and comparison of existing models. - Model coverage: Focuses on three prominent first-order normative frameworks; other ethical theories or hybrid approaches are not analyzed in depth. - Implementation specifics: Does not provide implementable algorithms or empirical tests for moral self-update mechanisms or bias mitigation in deployed systems. - Generalizability: Arguments about epistemic vs moral predictability and simulacrum risks are broadly reasoned; domain-specific constraints (e.g., medical, military) may yield varying practical outcomes not exhaustively examined.