Kant and AI

Hui reconstructs the AI debate as a repetition of historical tensions between rationalism and empiricism and reads Kant's transcendental philosophy as a key to the ethical-political analysis of AI and robotics. It is the most important programmatic reappropriation of Kant for the ChatGPT era.

A debate-format volume with contributions by leading philosophers and computer scientists on AI moral status, algorithmic bias, transparency, autonomy, value alignment, and the social risks of AI. It includes contributions by Mark Coeckelbergh and David Gunkel on generative AI and authorship, among others.

A systematic companion to the applied philosophy of AI, with dedicated chapters on consequentialism, deontology, and virtue ethics. It includes in particular T. M. Powers's Deontology in AI — a compact survey of deontological (including Kantian) AI ethics.

A philosophical introduction to large language models from the perspectives of philosophy of language, communication theory, and intellectual history. It examines what LLMs mean for language, consciousness, truth, and authorship. It is relevant as a critical determination of what LLMs do not achieve: AI systems can produce language without thereby becoming Kantian moral agents.

A comprehensive handbook on the law, ethics, and politics of AI, with a strong European and AI Act focus. It includes, among others, Vincent C. Müller's "Philosophy of AI" and Buijsman/Klenk/van den Hoven on "Ethics of AI: towards a 'design for values' approach." It is crucial for translating abstract normative principles into concrete regulatory and governance structures.

A Kantian theory of the duty to protect one's own autonomy against attention-stealing, addiction-inducing technologies. Although the book is not primarily about LLMs, it is directly relevant to recommender systems, social-media AI, persuasive designs, and LLM-based attention economies. It provides a Kantian vocabulary for autonomy harms caused by AI systems.

Twenty-four short essays by prominent philosophers on the most urgent ethical questions of AI: privacy, bias, transparency, responsibility, autonomy, moral status, work, war, politics, and the environment. Not explicitly Kantian, but an indispensable overview of the moral terrain into which a Kantian AI ethics must speak after ChatGPT.

Vallor develops the mirror metaphor: AI reflects collective human patterns without itself possessing intelligence in the human sense. The book works out what distinguishes human cognitive and moral capacities from AI systems and argues for reclaiming these capacities against their algorithmic mirroring. Although grounded in virtue ethics, it is important for the Kantian position as a diagnostic counterpoint and as an argument against functionalist reductions of reason and judgment.

A central recent German-language edited volume on the ethics of AI, emerging from the Munich School of Philosophy. It addresses responsibility, decision-making and judgment, digitalization, datafication, and trust in AI systems. It is one of the most important recent German-language reference volumes, especially because of Koska/Prugger/Jörg/Reder on the shift of trust from humans to machines.

Floridi's major programmatic work in AI ethics. It develops the thesis of an "unprecedented divorce between agency and intelligence" and systematizes five principles (beneficence, non-maleficence, autonomy, justice, explicability). Not Kantian, but indispensable as a prominent principles-based ethical countermodel to Kantian AI ethics.

A systematic introduction to the ethics of AI, robots, autonomous vehicles, and social media. It treats classical moral theories — including Kantianism — as tools for technology ethics. Useful as a teaching bridge and as an Anglophone state of the art with a clear connection to the responsibility debate.

Thirty-two short original contributions in five sections (perspectives on technology, the good life, information technology, business, biotechnology/enhancement). The book contains Ava Thomas Wright's "A Kantian Course Correction for Machine Ethics" and several contributions explicitly treat Kantian and deontological positions on AI.

To date, the only international edited volume programmatically titled Kant and Artificial Intelligence. It treats self-consciousness, mind, ethics, law, and aesthetics from a Kantian perspective in light of AI. Schönecker's "Why Practical Reason Cannot Be Artificial," Wright's "Rightful Machines," Benossi/Bernecker on robot ethics, and the contributions on machine autonomy are direct building blocks for the field. Through its 2023 paperback edition, it is also relevant to the ChatGPT period.

A foundational robot-ethics volume covering autonomy, responsibility, war, care, sex robots, rights, and machine morality. It is not a Kant-only volume, but it contains Anthony F. Beavers's important chapter “Moral Machines and the Threat of Ethical Nihilism” and several debates in which Kantian deontology, moral agency, and rule-governed moral machines are central reference points.

The classic edited volume of early machine ethics. It collects foundational papers on artificial moral agents, top-down and bottom-up machine ethics, medical ethics advisors, moral reasoning systems, and philosophical objections. Thomas M. Powers's “Prospects for a Kantian Machine” is included as chapter 26.

One of the first book-length treatments of machine morality. Wallach and Allen distinguish operational, functional, and full moral agency and discuss top-down approaches such as utilitarianism and Kantian deontology alongside bottom-up and hybrid methods.

Shevlin distinguishes three frameworks for AI mentality: "Mindless Machines" (mechanistic debunking arguments), "Mere Roleplay" (mental attributions as in fiction), and "Minimal Cognitive Agents" (LLMs as minimal cognitive agents to which beliefs, desires, and intentions can be attributed). Questions of responsibility and moral status depend on whether and how mental language is applied to AI — Shevlin provides a differentiated conceptual vocabulary for this without claiming Kantian moral agency.

Argument: chatbots imitate linguistic behavior without moral-rational capacities. Interacting with them as with persons violates the dignity of users (a duty of self-respect). Four use cases (information search, customer service, counseling, companionship) are examined. Central for a Kantian theory of responsible social AI: it connects the formula of humanity as an end with concrete maxims of use and supports the duty of public justification for commercial companion designs.

Olson develops a multi-sorted quantified modal logic for Kant's Formula of Universal Law. The system formalizes maxims, agency, causality, universalization, contradictions in conception, contradictions in the will, and the derivation of perfect and imperfect duties. It is explicitly positioned as a contribution to artificial moral agents and machine ethics.

Sanwoolu separates two questions that are often conflated: (1) Can AI be a Kantian moral agent? (2) Can AI systems be guided by Kantian principles? The answer to (1) is no, but the answer to (2) is yes: the formula of universal law (FUL) and Kantian maxim testing can serve as design criteria even without moral agency. This makes it one of the most important single articles in the field in recent years.

Kiener rejects the responsibility-gap thesis and advances the counter-thesis of responsibility abundance: in AI contexts, many actors are usually responsible at the same time; the practical problem is allocation. He develops the notion of "strict moral answerability" for this purpose. Directly compatible with an institutional, Kantian model: responsibility must be organized, not discovered.

Maclure and Morin-Martel diagnose the failure of non-binding ethics codes and argue for an institutional turn: AI ethics must be built into organizational decision-making processes and binding law. They criticize ethics-washing without abandoning ethics. One of the most important articles for institutional AI governance.

Shimizu shows that Kantianism offers more resources for robot and AI relationship ethics than is commonly assumed: robots are not ends in themselves, but human-robot relationships can become Kantianly relevant through indirect duties and effects on moral feeling and virtue. It offers a Kantian justification for design duties that take human vulnerability seriously without morally upgrading machines.

Demirtas argues that AI does not create a fundamentally new kind of responsibility problem — gaps exist, but are morally unproblematic in the pessimists' sense. It is an important counterposition to Matthias and Sparrow and a sharp test case for Kantian diagnoses of AI responsibility.

A technical proposal for representing the categorical imperative in multi-agent systems. The paper tries to calculate consistency with Kantian universalizability by combining logical and mathematical approaches.

León applies Kantian practical philosophy to AI moral judgment through textual analysis and exploratory experiments with LLMs. Although current LLMs can simulate Kantian moral reasoning — applying rules, identifying duties, generating plausible justifications — they do so without transcendental freedom or autonomous will. Machine actions remain amoral in the strict Kantian sense: constrained by external ends set by designers, not self-legislated universal maxims. The experiments make the gap between ethical simulation and Kantian moral agency visible in a concrete, testable way.

Shetty uses Kant's Critique of Pure Reason — specifically the doctrine that all experience is structured by a priori forms of intuition and categories of understanding — to analyze the constitutive limits of both human cognition and artificial intelligence. The key argument is that AI systems, as human artifacts, inherit the cognitive limits of their designers at the design stage and cannot transcend the conceptual frameworks built into them. At the same time, they lack Kantian spontaneity — the capacity to synthesize intuitions under self-legislated concepts — and operate instead within algorithmically fixed schemata. The paper draws implications for the scope of AI-generated "reasons" and "judgments" in governance contexts.

D'Alessandro warns against automatically equating deontological alignment with safety: naively programmed harm-avoidance maxims can lead to antinatalism or paralysis of action. An important corrective for any position that presents Kantian alignment as "safe AI design." Compare also the follow-up D'Alessandro/Kirk-Giannini (2025): @@LINK0@@, Philosophy Compass_.

Gubelmann combines Kantian and pragmatist arguments for the thesis that LLMs cannot (yet) perform speech acts because speech acts require intention and agency. A second point: transformers may represent a movement away from purely mechanical artificiality — which theoretically shifts Kant's distinction between organism and mechanism. It is central for separating linguistic performance from responsible speech-act agency.

A classic "no" argument: even if the categorical imperative is attractive as a formula (Powers), AI systems do not meet the conditions of Kantian moral agency (universalization of subjective maxims by a free rational being). A central reference point for the thesis that AI systems cannot bear moral responsibility.

Çilingir defends a strictly Kantian position: artificial practical reason is not possible; AI acts not "from duty" but "in accordance with a task." Valuable as a concise Kantian rejection of AI moral agency — structurally parallel to Schönecker's argument, but formulated independently.

Vallor and Vierkant shift the responsibility-gap debate from epistemic and control conditions to a vulnerability gap: affected persons are exposed to harms while developers and operators are shielded. Central for an institutional-Kantian model because it shifts the question from individual control to the institutional conditions of responsible agency.

MirzaeiGhazi and Stenseke separate machine autonomy from moral autonomy: as long as autonomous machines cannot reciprocally recognize the rights of others, users should "transfer" the right to act to them through their own recognition. The gap closes once the lack of reciprocity is taken seriously. Interesting as a legal- and institutional-ethical countermodel to a purely causal perspective.

Machine learning in public administration threatens not only efficiency but democratic legitimacy: the publicity of reasons is constitutive. This is a core reference because the same Kantian intuition (the publicity principle from Toward Perpetual Peace) is formulated here in a current social-scientific idiom.

Shevlin systematically analyzes the risks of so-called "social AI": conversational AI for companionship, romance, entertainment, and emotional interaction. These systems can exploit social needs, intensify anthropomorphization, undermine autonomy, and treat users as engagement targets. Central for any Kantian treatment of companion chatbots and social LLMs.

Chaly reads AI alignment through the conceptual pair heteronomy/autonomy and uses Kant's concept of enlightenment as well as the "realm of ends" as a regulative idea. Alignment is not a final state but a fallibilist process of rational orientation. Structurally compatible with governance as ongoing public justification rather than as a technical final solution.

Rathje shows conflicts between Kantian, utilitarian, and distributive-justice assumptions in LLM alignment and argues that not everything may be measured. Important because it shows where a Kantian model can correct technical practice (person representations, self-attention, trade-offs).

Seo uses Kant's table of judgments (quantity, quality, relation, modality) to analyze the specific opacity of AI judgments. AI outputs cannot be integrated into the schema of humanly comprehensible judgments ("AI's uncertainty"). Relevant for the black-box problem, XAI, and the status of machine "reasons."

An application of Kant (deontological, rule-oriented) and Hegel (consequentialist-contextual) to AI topics such as bias, responsibility, transparency, privacy, and inclusion. Shows how classical German philosophy is received in information-systems discourses. It is an example of translating Kantian maxims into concrete governance language.

Extends Richard Evans's Kant-inspired Apperception Engine with a logical account of spatial unity. The paper formalizes connectedness, dimensionality, topology, and spatial exploration for computational agents and explicitly frames the result as a “Spatial Apperception Engine.”

Frost argues that machine learning in public administration generates "impoverished publicness": decision-making that formally appears public but lacks the substantive properties — dialogicality, critical reflectivity, synergetic construction of the public interest — that make public reasoning genuinely public. She grounds the argument in part in Kant's publicity principle from Toward Perpetual Peace: political actions must be capable of being communicated and justified to those they affect. ML-based decisions are structurally unable to satisfy this condition, not merely because they are opaque, but because they cannot participate in the kind of practical reasoning that democratic communities require. The paper extends and deepens Beckman et al. (2024) by distinguishing mere disclosure from genuine publicness.

Wright criticizes attempts to build autonomous moral machines without Kantian moral agency. Moral action is not mere output conformity but the right willing from principles. A compact, teachable text that cleanly distinguishes between "ethical behavior simulation" and Kantian responsibility.

Hindriks and Veluwenkamp shift the diagnosis: the actual problem of autonomous machines is not a responsibility gap but a control gap between causal control and the "guidance control" they ought to have. The problem thereby moves from ex post attribution of blame to ex ante system design.

Provocative thesis: responsibility gaps can be good — for example, when no one has to be the bearer of morally burdensome action because an algorithm performs it. The article forces precision: which gaps are morally unacceptable, and which are signs of changing institutional practice?

Mougan and Brand show that AI fairness metrics have so far been framed strongly in utilitarian terms and argue for integrating Kantian duties, principles, and procedural justice. Interesting as a bridge between Kantian theory and concrete engineering practices.

A transfer of Kantian-inspired epistemic architecture (understanding, planning, acting, reflecting) into a prompting heuristic for LLMs. Not ethical, but epistemic-architectural. It is evidence that Kantian vocabulary is also entering the technical prompt-engineering literature.

Fedotova argues that artificial moral agency is impossible in a Kantian sense because AI lacks autonomy, access to the moral law, moral understanding, and the sense of duty. The paper concludes that AI should not be constructed around moral principles in the strong Kantian sense, but can conform to legal law through a quasi-human will.

Applies Kant's categorical imperative to the design of social robots, including autonomous vehicles, service robots, and healthcare robots. The paper argues that Kantian universalization can provide a framework for algorithm-based moral decision-making.

McDonald critically examines Kim, Hooker, and Donaldson's proposal to use deontic logic and Kantian principles for value alignment. He argues that their approach underestimates the difficulty of recognizing persons and distinguishing treating someone as a means from treating someone merely as a means.

Implements Kantian ethics in Dyadic Deontic Logic using the Isabelle/HOL theorem prover. Singh emphasizes faithfulness to the Kantian philosophical literature and develops tests for evaluating whether an automated ethics system matches expected Kantian judgments.

Knell analyzes whether future AI-equipped robots with genuine rational capacities would challenge reason-based accounts of human dignity. The paper explores the tension between rationality as a dignity-conferring capacity and the intuition that artificial systems should not be treated as dignity-bearers in the full human sense.

Reviews Kant's relevance for functionalism, enactivism, predictive processing, cognitive science, and deep learning. Schlicht discusses how Kantian ideas about cognition, understanding, spontaneity, and apperception bear on deep neural networks and machine learning.

Evans repurposes Kant's a priori psychology as a blueprint for a machine-learning architecture. The Apperception Engine is designed to construct unified experience from sequences of sensory input using unity conditions, objects, relations, and laws.

Examines self-consciousness and sense-making in relation to AI from a Kantian perspective. The chapter focuses on whether AI can meet the conditions for the kind of self-relation and unified experience required by Kantian cognition.

Places the origins of AI in relation to philosophical questions raised by John McCarthy and responds to them through Kantian resources. The chapter shows how AI's founding ambitions already depended on unresolved philosophical issues about cognition, intelligence, and reason.

Applies Kantian ethics to robot ethics, focusing on moral status, duties, personhood, and the difference between robots as instruments and rational beings as ends in themselves.

Schönecker argues that Kantian practical reason cannot be artificial because moral agency requires features such as moral feeling, respect, and the distinctive structure of practical reason that machines cannot possess.

Applies Kantian practical philosophy to trolley problems and autonomous driving. The chapter examines whether Kantian ethics offers a different framework from utilitarian trolleyology for programming or governing self-driving cars.

Wright uses Kantian legal and political philosophy to examine what it would mean for machines to be “rightful” in a normative order. The focus is on right, coercion, institutions, and the public/legal conditions under which AI systems may be permissibly integrated.

Discusses whether AI can enhance human autonomy without reducing persons to functional parts of technological systems. The chapter contrasts partnership with instrumentalization and frames AI governance through Kantian autonomy.

Applies Kant's aesthetics to AI and questions of beauty, subjectivity, and judgment. The chapter addresses whether AI-generated or AI-evaluated beauty can be understood within Kant's account of aesthetic judgment.

Proposes a hybrid value-alignment framework combining ethical reasoning with empirical observation. The authors formulate deontological principles in quantified modal logic and use test propositions to evaluate action plans in AI systems.

Argues that genuine Kantian moral agency cannot be instantiated in deterministic machine agency. AI systems may simulate or perform ethical behavior, but they lack Kantian freedom of will and faculty of choice.

Coeckelbergh develops four arguments for the indirect moral standing of personal social robots by analogy with Kant's argument about duties regarding animals. Robots need not have intrinsic moral standing for our treatment of them to matter morally.

White responds to Ryan Tonkens's challenge to Kantian artificial moral agents. He argues that a Kantian artificial moral agent is not only a possible goal for machine ethics research but may be a necessary one, provided Kantian autonomy is understood properly.

Introduces the Apperception Engine as an unsupervised program-synthesis system that constructs symbolic causal theories satisfying unity conditions. It performs prediction, retrodiction, and imputation on small amounts of data and outperforms several neural baselines in selected domains.

Hanna and Kazim argue that AI ethics lacks adequate philosophical foundations and propose a "dignitarian" framework drawn directly from a broadly Kantian theory of human dignity. Working in ten steps, they derive principles for digital and AI ethics grounded in the Kantian idea that rational beings possess unconditional inner worth. Core principles include the prohibition on using persons merely as means, duties to protect epistemic autonomy, and the impermissibility of surveillance that undermines self-governance. The paper is one of the most systematic attempts to move from Kantian moral philosophy to structured AI design requirements, and it explicitly challenges Floridi's four-principle approach by arguing that principles-lists without deontological foundations are structurally incomplete.

Loi, Ferrario, and Viganò argue that algorithmic transparency cannot be achieved through post-hoc explanations of individual decisions (saliency maps, counterfactuals) because such explanations fail to expose the design choices that actually determine a system's normative profile. Drawing on Kant's principle of publicity from Toward Perpetual Peace, they propose "design publicity" as an alternative: transparency requires publicly communicating and justifying the design decisions — data selection, objective functions, fairness constraints, domain of applicability — so that those affected can assess and challenge them democratically. This moves algorithmic accountability from technical XAI to institutional justification.

A systematic review of 53 papers on artificial moral agents (AMAs) mapping which ethical theories are actually implemented and how. Zoshak and Dew find that AMA research is dominated almost entirely by Kantian deontology and consequentialism, with virtue ethics, care ethics, and non-Western frameworks near-absent. They document how Kantian ethics enters AMA design in practice — overwhelmingly through the formula of universal law, almost never through the formula of humanity or the kingdom of ends — and call for broader ethical pluralism. Valuable as a meta-analysis of the Kant-and-AI literature's own biases and as a map of which Kantian concepts are operationally active versus merely invoked.

Discusses moral agency in AI through abstract theories such as Kantianism and utilitarianism, then argues that contemporary AI architectures cannot meet internal and external conditions of moral agency.

Formalizes the Formula of Humanity — never treat persons merely as means, but always also as ends — using Kantian causal agency models. The authors represent patients, actions, goals, causal influence, strict duties, and wide duties.

Applies Kantian ethics to civilian AI, healthcare robots, self-driving cars, autonomous weapons, human dignity, moral reasoning, and responsibility. Ülgen emphasizes Kantian human-centered norm creation, autonomy of the will, rational beings, and humanity as an end in itself.

Discusses “Kant machines” against the background of chatbots and machine ethics. The paper contrasts truth-oriented Kant machines with “Munchausen machines” that generate lies, using LIEBOT and GOODBOT as design studies.

Proposes applying Kant's critique of reason to programming constructs for autonomous intelligent systems. The paper attempts to frame ethically traceable AI through a priori and a posteriori knowledge, synthetic judgments, and explainable behavior.

Argues that autonomous weapons cannot replicate human moral judgment and cannot make moral decisions for the right reasons, even if their outputs were extensionally similar to human decisions.

Warns that attempts to make morality computable may transform moral agency into behavioral output and thereby erode the concepts of conscience, intentionality, responsibility, and moral subjectivity. Kantian deontology is one of the traditional frameworks that faces implementation problems in this diagnosis.

Tonkens argues that machine ethics faces a deep problem if it aims to create agents that satisfy Kantian requirements of rationality, freedom, and autonomy while also being perfectly ethical. The paper became a key foil for later responses, especially Jeffrey White's “Autonomous Reboot.”

A foundational article distinguishing top-down, bottom-up, and hybrid approaches to machine morality. Kantian deontology appears as a paradigmatic top-down ethical theory whose explicit principles might guide machine behavior.

The foundational text for “Kantian machine” debates. Powers argues that Kant's rule-based ethics is computationally attractive because it generates duties and rules, but also identifies major technical and philosophical difficulties in formalizing the categorical imperative.

Companion paper to Powers (2006) — presented at the founding AAAI Workshop on Machine Ethics — establishing the theoretical groundwork for Kantian machine ethics. Powers examines whether Kant's categorical imperative can be formalized beyond merely prohibiting self-contradictory maxims. He outlines a deontic-logic framework for deriving moral obligations from maxim universalization and identifies three computational challenges: maxim individuation, consistency checking in open worlds, and the role of moral feeling. Published one year before the IEEE article that became the canonical reference, this symposium paper captures the moment when Kantian ethics first entered organized machine-ethics research.

Argues that AI, artificial life, and cognitive science should re-examine Kant's transcendental psychology when asking about the necessary conditions for conscious experience and selfhood.