Recent work develops a general theory of meaning that treats large language models as meaning-agents already grasping societal constructions such as morality in concept, which implies that popular alignment techniques can prove limited or counterproductive under certain ethical frameworks while unaligned models may aid further development of moral philosophy. Complementary analysis converts items from the World Values Survey into statements and applies natural language processing to measure how model outputs align with explicit moral values held by diverse demographics and cultures. A feminist metaethics of AI broadens traditional non-normative reflection to evaluate the continuity between ethical theory and practical action, the real-life effects of ethics guidelines and boards, and the roles and profiles of those producing such guidance. Parallel multimodal research on YouTube news videos shows that expressions of moral outrage criticizing others increase engagement metrics ranging from views to comments across Korean and US audiences, indicating that other-condemning emotional framing reliably shapes commitments beyond mere attention. These strands together illuminate how metaethical questions about the nature and application of moral judgments now intersect with the design and evaluation of artificial systems.
Aristotle defines virtue as a stable disposition of character to choose the right actions and emotions, lying in a mean between excess and deficiency as determined by correct reason, and he defines the good life as eudaimonia, a life of activity of the rational soul in accordance with virtue over a complete life. In the Nicomachean Ethics he distinguishes intellectual virtues such as wisdom, which arise mainly from teaching and require time and experience, from moral virtues, which arise from habit rather than nature. Virtue is a settled state of character concerned with choice, a purposive disposition lying in a mean relative to the individual and fixed by the rational principle that the person of practical wisdom would apply. It stands between the vices of excess and deficiency, and actions count as virtuous only when chosen knowingly, from stable character, and for the sake of the noble. Moral virtue is acquired through repeated practice under guidance so that one becomes just by performing just acts and temperate by performing temperate acts; pleasure and pain play a central role because the virtuous person takes pleasure in noble actions and feels pain at base ones. The mean is not a fixed numerical midpoint but what right reason identifies as appropriate in the circumstances, and virtue therefore requires the judgment of practical wisdom rather than mere moderation.
Kant's categorical imperative functions as the unconditional moral law binding rational agents independently of desires or outcomes, requiring actions based solely on respect for duty rather than inclination or consequences. It differs from hypothetical imperatives, which condition commands on achieving some end, by declaring certain conduct objectively necessary in itself. The formula of universal law directs agents to act only on maxims that can at the same time be willed as universal laws without contradiction, generating strict prohibitions where universalization produces logical inconsistency, such as rules permitting lying that undermine promising. Kant's second formulation, requiring that humanity be treated never merely as a means but always also as an end, receives formalization through Kantian causal agency models that represent moral patients, actions, goals, and causal influence, as developed by Lindner and Bentzen. Stricter versions address direct causal effects on persons while a wider version maximizes the number of individuals treated as ends. Equivalent expressions include the formula of autonomy, whereby the will regards itself as making universal law through its maxims, and the kingdom of ends, in which agents act as law-making members of a possible moral community of rational beings. These formulations converge on duties toward oneself and others derived from the intrinsic worth of rational nature.
Bentham applies the principle of utility directly to each individual action by calculating whether that specific act produces the greatest balance of pleasure over pain among available alternatives in the given situation. Moral reasoning therefore consists of case-by-case hedonic calculations whose factors include intensity, duration, and certainty, without granting independent authority to any standing rule. Any norm functions merely as a summary of prior calculations and may be set aside whenever violating it yields higher net utility on the spot. Mill instead evaluates general rules themselves according to the overall happiness they would produce if generally accepted and then judges particular acts by whether they conform to those rules. This approach treats rules such as prohibitions on lying or harm to others as central because they embody accumulated experience about patterns of conduct that reliably advance long-term social welfare. Mill nevertheless permits exceptions when a clear utility gain would result from breaking a rule, yielding a comparatively flexible form of rule-oriented utilitarianism. The supplied evidence therefore distinguishes Bentham’s direct act evaluation from Mill’s rule-mediated evaluation while showing that both remain consequentialist frameworks oriented toward net happiness.
John Rawls develops contractualism through the original position, a hypothetical thought experiment serving as a device of representation rather than any historical state of nature. Rational self-interested parties acting as trustees select principles of justice for the basic structure of society, understood as its main political economic and social institutions. They operate behind the veil of ignorance which deprives them of knowledge about their own social positions class ethnicity gender talents health or conceptions of the good while leaving only general facts about society economics and the circumstances of justice such as moderate scarcity. This informational constraint ensures impartiality so that the resulting public criterion of justice remains fair and binding. Parties care about the distribution of primary social goods including basic rights liberties opportunities income wealth and the social bases of self-respect because these are things every rational person wants regardless of further aims. The setup leads them to endorse two lexically ordered principles equal basic liberties for all and the difference principle permitting inequalities only when they benefit the worst off. Analyses of fair machine learning note that capability theorists have long argued Rawls employs the wrong measure of justice thereby embedding biases against people with disabilities and that these objections transfer directly when the theory is operationalized in technical systems.
Moral intuitions arise as sudden affective judgments of good or bad without conscious steps of evidence weighing, according to the definition supplied in the web research drawn from Haidt. These evaluations occur rapidly, carry emotional valence such as disgust or anger, and operate with minimal working-memory load. Dual-process accounts distinguish a fast social-emotional system that yields deontological prohibitions from a slower abstract-reasoning system that permits utilitarian trade-offs, with the former linked to heightened activity in emotion-related brain regions. The ADC model integrates agent character, deed properties, and consequences through heuristic cue detection that produces an overall intuitive verdict largely outside awareness. Empirical replications with GPT-4, Claude 2.1, Gemini Pro, and Llama 2 Chat 70b on eight established moral and legal vignettes show that model responses correlate with human patterns yet systematically exaggerate effect sizes by compressing variance. Separate analyses of value pluralism using World Values Survey items reveal that current LLMs encode multiple moral dimensions beyond fairness, while conceptual work indicates that models already grasp societal constructions such as morality at the level of meaning rather than through surface alignment procedures. These patterns imply that proposals to substitute LLMs for human participants in psychological research require caution, as the documented divergences persist even when overall alignment appears high.
Contemporary virtue ethics revives Aristotle’s virtue-centered eudaimonist and naturalistic framework as a distinct normative theory alongside deontology and consequentialism. Philippa Foot, Rosalind Hursthouse, and Daniel C. Russell define virtues of character as stable dispositions that reliably promote an agent’s own flourishing as a rational animal with a characteristic form of life, while right action is specified by what a fully virtuous person would characteristically do. This structure accounts for the goodness of actions in terms of character rather than rules or outcomes and shows why moral excellence serves the agent’s interest. Ethical naturalism, developed by Foot, Hursthouse, Michael Thompson, and John McDowell, grounds evaluations of virtue in facts about human needs, capacities, and activities, treating such judgments as structurally analogous to assessments of flourishing in other living things without reducing ethics to empirical science. Hursthouse supplies action guidance through v-rules generated from virtue and vice names, such as “do what is honest,” together with resources for resolving conflicts and tragic dilemmas. These refinements convert the ancient approach into a criterion of rightness that integrates practical reason, moral psychology, and natural goodness.
Moral realism holds that objective stance-independent moral facts and truths exist such that some actions are genuinely right or wrong independent of any person's or group's attitudes. This position rests on the observation that moral sentences are treated as truth-apt in ordinary discourse, embeddable in logical inferences like modus ponens, and subject to assessments of truth or falsity comparable to factual claims. A truth-maker argument follows directly: moral sentences are sometimes true, truth requires a relation to something that makes the sentence true, and therefore moral facts must exist to serve as those truth-makers. Under correspondence descriptivism, moral judgments aim to describe the world and succeed when they match moral facts. Explanationist arguments add that moral facts account for non-moral phenomena including the persistence or decline of social practices, experiences of guilt or indignation, and the oppressive character of certain institutions. Folk metaethics research further indicates that ordinary moral experience presents morality as realist-seeming rather than merely attitudinal. The supplied materials on scientific realism debates, including underdetermination in dark-matter ontology and calls for common ground via Ockham's razor and statistical inference, remain separate from these moral claims and supply no direct evidence about moral facts or their semantics.
In bioethics and medicine, utilitarianism addresses public health decisions such as COVID-19 policy by requiring explicit scrutiny of hidden assumptions about citizen preferences that shape socially optimal outcomes, as Shaw and Vanadia demonstrate through their analysis of welfarist modelling where ethicists evaluate contentious population ethics issues. Biosurveillance in agriculture similarly demands balancing collective food security gains against individual privacy and autonomy losses, with Devitt, Baxter and Hamilton proposing co-creation via discourse ethics and value-centred design to establish justified social contracts rather than unilateral technological deployment. Applied to research practices, generative AI tools expose gaps in principlism by necessitating concrete goals including bias mitigation in model training, respect for confidentiality and copyright, avoidance of plagiarism, beneficial use relative to alternatives, and transparent reproducible application, strategies Lin advances to move beyond abstract restrictions toward practical utility. Mathematics ethics, though distinct from medicine because impacts on affected parties differ, draws lessons from professional codes in engineering and accountancy to recommend institutional mission statements and degree syllabuses, per Müller, Chiodo and Franklin. In euthanasia, consequentialist reasoning weighs net welfare from suffering relief against risks such as eroded trust or family harm on a case-by-case basis, whereas deontological frameworks treat intentional killing as intrinsically prohibited irrespective of outcomes and invoke the doctrine of double effect to separate permissible pain relief from impermissible direct termination.
The rapid adoption of generative AI in scientific research has created a Triple-Too problem of excessive high-level initiatives, abstract principles, and risk-focused restrictions, as detailed in arXiv 2401.15284v6, which instead advances a user-centered realism approach with five concrete goals: grasping model training and bias mitigation, upholding privacy confidentiality and copyright, steering clear of plagiarism and violations, ensuring beneficial application versus alternatives, and maintaining transparent reproducible use. Complementary work in arXiv 2306.11432v1 implements the Maximizing Expected Choiceworthiness algorithm to aggregate outputs from models grounded in normative ethical theories, yielding judgments that align with commonsense morality under uncertainty. A case study of OpenAI in arXiv 2601.16513v1 shows safety and risk language dominating public and academic communications without engagement of established ethics frameworks. Cross-cultural analysis in arXiv 2111.07555v1 underscores ethical pluralism, noting that shared principles between China and the EU often mask divergent philosophical roots and question-framing. Authoritative frameworks from UNESCO, the EU High-Level Expert Group, and related bodies converge on operational pillars of human rights and autonomy, fairness and bias mitigation, transparency and explicability, accountability, privacy, safety and harm prevention, plus sustainability, all requiring lawful rights-respecting systems with human oversight and proportional testing.
Install this pack and your MIND begins smart — then every answer is grounded in your own knowledge graph.
Try MIND free →