← All sources

Kantian deontology for AI: alignment without moral agency

Oluwaseun Damilola Sanwoolu · 2025 · AI and Ethics 5:5425-5437   interlocutor high priority coded

Main argument

Thesis: AI can be aligned with Kantian deontology WITHOUT being a Kantian moral agent - alignment requires only that AI be morally CONSTRAINED (act in accordance with the Formula of Universal Law via simulation), not morally ACCOUNTABLE (act from duty, praise/blame). Argument type: conceptual, Kant exegesis + application. Two objections answered: (1) AI lacks Kantian moral agency (free will, rational autonomy, self-legislation) - Sanwoolu ACCEPTS this (AI outputs are heteronomous/probabilistic, not autonomous) but argues, via the child-learning-honesty and industrial-software/sniffer-dog/rock analogies, that morally good action can occur without moral agency, and that having morally-significant consequences does not make something a moral agent; (2) the particularist challenge (Dancy/Anscombe: the CI tests maxims but can't form them or capture context) - answered via Herman's rules of moral salience and Bremner/Dunn's reading of Kantian practical/reflective judgment, with the claim that transformer models are a functional equivalent enabling context-sensitive maxim formation. Chooses the Formula of Universal Law specifically because the other CI formulas presuppose will/rational-humanity that AI lacks.

Why it matters here

The closest published analogue to Augustine's own position: accepts that AI is NOT a moral agent (the anti-Luke conclusion) yet argues normative principles still apply to it - via the constrained/simulation route. Directly models the exact argumentative structure Augustine needs (concede no moral agency, retain normative evaluability) and, being by a Nigerian-affiliated scholar at Kansas, is a natural interlocutor. Also a rare defense of a SINGLE framework (Kant) as alignment target, useful as the foil for convergentism.

Reading notes

Full close read completed. 13pp. Oluwaseun Damilola Sanwoolu, University of Kansas (Nigerian name - potentially relevant to Augustine's African-philosophy positioning and network). Explicitly engages Gabriel 2020, Moor's moral-agent hierarchy, and the particularist (Dancy/Anscombe) challenge. Distinctive claim: transformer models are the 'functional equivalent' of Kantian practical judgment for maxim-formation - a contestable move worth flagging.

Sanwoolu, O. D. (2025). Kantian deontology for AI: alignment without moral agency. AI and Ethics, 5, 5425-5437. https://doi.org/10.1007/s43681-025-00784-8

Close reading — 13 coded units

#1 · pp. 5425 · claim
“AI alignment with Kantian principles does not require moral agency in Kant's sense. I propose that the Categorical Imperative (CI) can serve as a useful framework for AI alignment, guiding the creation of maxims governing AI actions and testing their universalizability, particularly using the first principle of the CI which is the formula of the universal law (FUL).”
#2 · pp. 5426 · argument
“Some philosophers have argued that we ought not align AI systems with any ethical theory as there is an absence of moral agreement. Since there are numerous ethical positions, how do we decide which ethical theory to turn to? [...] I reject the idea that the multiplicity of ethical theories provides sufficient justification for bypassing ethical theories as candidates for AI alignment.”
#3 · pp. 5426 · argument
“the lack of consensus on the correct ethical theory is not sufficient grounds to bypass ethical theories in this discussion. Concerning human morality, for example, one could adopt pluralist ethical approaches. Although ethical theories all have strengths and limitations, yet we continue to seek moral guidance despite the lack of agreement on a single 'correct' theory.”
#4 · pp. 5428 · argument
“if we were to encode Kantian duties or laws into an AI system, it would be compelled to follow these laws, as that is how it is designed to operate. This circumstance contradicts the principle of rational autonomy. Typically, machine outputs are probabilistic, meaning that they are heteronomous in nature rather than autonomous; their source of law is from external programming rather than self-determination.”
#5 · pp. 5428–5429 · argument
“AI systems, however, lack both aspects [of freedom]. They are constrained by pre-programmed architectures, optimization objectives, and statistical learning from data. Consequently, they do not exhibit negative freedom, as they cannot truly deviate from causal determination, nor do they exhibit positive freedom, as they cannot will or legislate moral law from reason. [...] I argue that we refrain from ascribing moral agency to AI, particularly within the Kantian ethical framework.”
#6 · pp. 5429 · argument
“a morally good action can occur without full moral agency. This distinction is particularly relevant when considering artificial intelligence. AI systems can be designed to act in accordance with moral guidelines without possessing the capacity for independent moral reasoning.”
#7 · pp. 5429 · argument
“Is it necessary to view all agents whose actions have moral consequences as moral agents? [...] we would be wrong to assume that all entities or systems whose actions have normative or moral outcomes are moral agents. [...] if a banking software accidentally overpays someone, it has moral consequences, yet we don't attribute moral judgment or agency to the software itself. Similarly, if a sniffer dog fails to detect illegal drugs [...] we still wouldn't regard the dog's actions as worthy of praise or blame.”
#8 · pp. 5429 · argument
“morality is contingent on the nature of the entity performing the action. Humans, unlike AI or non-living objects, possess the capacity for moral deliberation and autonomy. Therefore, the fact that AI systems may generate actions with moral consequences is not enough to classify them as moral agents. And we can still take the consequences of their actions seriously without ascribing moral agency to them.”
#9 · pp. 5429 · argument
“some scholars such as Talbot et. al argue that AI systems should not be held to deontological standards at all, on the grounds that deontology, unlike consequentialism, requires moral agency. [...] I agree with the diagnosis that AI systems are not moral agents [...] However, I reject the prescription that follows namely, that this rules out the application of deontological principles to AI design and behavior.”
#10 · pp. 5430 · claim
“while AI systems cannot be Kantian agents, they can still be guided by deontological principles, particularly through simulation. [...] The key shift is from holding AI morally accountable to holding them morally constrained. AI can simulate the process of acting on maxims and testing them for universalizability, functionally resembling Kantian deliberation even if they do not act from duty.”
#11 · pp. 5430 · objection
“Dancy and Anscombe maintain that the CI cannot tell us how to formulate maxims. They argue that the CI can only tell us how to determine the moral permissibility of maxims that have already been formulated, however, if we want AI systems to apply the FUL, we want it to know how maxims are formed.”
#12 · pp. 5430–5431 · argument
“Herman maintains that 'to be a moral agent one must be trained to perceive situations in terms of their morally significant features (as described by the RMS)'. [...] the teachability of RMS suggests that AI could potentially learn to identify morally salient features in situations.”
#13 · pp. 5433–5434 · claim
“[AI has] a functionally equivalent mechanism—transformer models—which can allow them form maxims that consider morally salient facts. Thus, supporting the claim that AI alignment is possible within a Kantian framework.”

Synthesis-matrix row

complicates T4-ROSSIAN-DEMAND
retains single framework (Kant) against the proceduralist retreat
supports T5-AGENCY-DENIED-EVALUABILITY-KEPT
constrained-not-accountable argued from Kant

Memos (5)

thesis-link · unit #7
This paper is the published mirror of Augustine's own position and gives him both an ally and a sparring partner. THE ALLY: Sanwoolu independently reaches Augustine's anti-Luke conclusion - AI is not a moral agent (units 5,7,8) - and, crucially, provides the exact rebuttal to Luke's inference (from AI-has-consequences/rationality to AI-is-responsible): units 7-8, the banking-software / sniffer-dog / rock cases show having morally-significant consequences does NOT entail moral agency. Augustine can cite Sanwoolu as independent published support for the experiment's conclusion, which directly answers Howard's 'unnecessary territory' worry - a peer-reviewed 2025 paper builds its whole argument on establishing exactly this point. THE DIFFERENCE: Sanwoolu argues from armchair Kant exegesis; Augustine argues from an empirical LLM experiment. Frame the experiment as the empirical complement to Sanwoolu's conceptual argument.
theoretical · unit #10
The constrained-vs-accountable distinction (AG-CONSTRAINED, unit 10) is a sharp conceptual tool the whole dissertation can adopt: it cleanly separates two questions the literature conflates - (a) is the AI subject to/guided by norms? (morally constrained - yes) and (b) does the AI bear responsibility for norm-violations? (morally accountable - no). This maps precisely onto Gabriel & Keeling's stipulation (GABRIEL_KEELING unit 10: 'morally evaluable actions' without 'moral claims') - Sanwoolu ARGUES what G&K ASSUME. And it sets up the dissertation's central move: if AI is constrained-but-not-accountable, then responsibility for its norm-violations MUST land on humans/institutions (RL-DEV/USER/INST), never RL-SYS - which is exactly why the responsibility-attribution question (Kästner) is the unavoidable next step the alignment literature keeps bracketing.
comparison · unit #3
Interesting divergence from the field's dominant response to pluralism. Faced with 'no single true theory' (Gabriel 2020 unit 15), the field mostly RETREATS from first-order theory to procedure (Gabriel proceduralism, Zhi-Xuan/G&K contractualism). Sanwoolu (units 2-3) REFUSES that retreat: he explicitly invokes pluralist ethical approaches as precedent for keeping first-order theory (Kant) despite disagreement. This is structurally the dissertation's own move - keep object-level normative theory (Rossian pluralism), don't retreat to bare procedure. Sanwoolu is thus a methodological ALLY against proceduralism even though he backs a single framework (Kant) where the dissertation backs convergent pluralism. Stage the lit review as: proceduralists (Gabriel, Zhi-Xuan, G&K, Schuster-Kilov) vs framework-retainers (Sanwoolu, and the dissertation) - a cleaner axis than pro/anti-preferentism.
theoretical · unit #13
Flag for critical engagement: Sanwoolu's load-bearing claim that transformer models are the 'functional equivalent' of Kantian practical judgment (units 12-13) sits in direct tension with MILLIERE_2025 and LI_2026's Khamassi strong/weak distinction. Millière shows LLMs do NOT reliably perform contextual normative weighing (they follow disposition-activation, not salience-tracking); Khamassi shows LLMs pass pattern-matching but fail intentionality/causal reasoning. So Sanwoolu's functional-equivalence claim is empirically contestable by two other coded sources. This is a live disagreement the dissertation can adjudicate WITH DATA: Augustine's LLM moral-reasoning experiment directly tests whether transformer 'maxim formation' tracks morally salient facts or just salient prompt features. If Millière is right, Sanwoolu's Kantian-alignment proposal fails at exactly the practical-judgment step - and Augustine has the experiment to show it.
thesis-link
Network/positioning note: Oluwaseun Damilola Sanwoolu (Yoruba name, U Kansas) is a natural node for Augustine's African-philosophy positioning and scholarly network - a philosopher of African origin publishing on AI ethics + Kant. Combined with the Ubuntu concessions in GABRIEL_2020 (fn14) and SCHUSTER_KILOV (fn1), and Metz's African moral theory (cited by Gabriel), there is a coherent 'African philosophy and AI alignment' thread forming across the corpus worth its own lit-review subsection and possibly worth reaching out to as part of building the dissertation's scholarly community.