Thomistic Natural Law Theory Applied to AI Agent Decision-Making

Thesis

Thomistic natural law is not fundamentally a rulebook. It is an architecture of practical normativity: rational creatures participate in eternal law; they possess first practical principles through synderesis; they derive more particular norms through reason; they apply those norms to cases through conscience and, more fully, through prudence; and their practical reasoning culminates not merely in propositions but in choice and action. That layered picture maps surprisingly well onto contemporary AI governance stacks—system instructions, constitutions, policy layers, and case-specific reasoning—but only as a functional analogy. In Aquinas, these layers presuppose intellect, will, appetite, and an intrinsic teleology toward the good. Current language-model agents are instead externally steered, probabilistic systems that generate tokens from learned distributions under platform-, developer-, and user-level constraints. Thomism is therefore more illuminating as a framework for AI governance and system design than as a basis for attributing literal prudence, virtue, or moral agency to present models. (New Advent)

A second thesis follows. Aquinas gives a sharper vocabulary than most alignment discourse for distinguishing four things that are often collapsed together: the source of norms, the derivation of more specific rules, the application of general norms to singular cases, and the motive power by which a judgment becomes an action. AI engineering usually has decent names for the first three and weak ones for the fourth. Thomism explains why that last gap matters: in a full practical act, reason does not stand alone. Judgment is followed by choice, and choice belongs to the will. That point is where the analogy with AI becomes most illuminating—and most fragile. (New Advent)

1. Aquinas's framework of natural law

1.1 Law, eternal law, and natural law

Aquinas begins with a very strong claim: law is "a rule and measure of acts," and belongs to reason because reason directs action to an end. He adds that law is properly an ordinance of reason for the common good and that it binds only if it is promulgated. Natural law is not an exception to this framework. It is law precisely because it is the rational creature's participation in eternal law, and it is promulgated because God has "instilled it into man's mind" so that it is naturally known (ST I-II q.90 a.1, a.4; q.91 a.2). This already matters for AI analogy. Natural law, for Aquinas, is not merely external command. It is an intelligible order internal to rational creatures as participants in divine providence. (New Advent)

That participation thesis prevents a common misunderstanding. Natural law is not simply "whatever humans are inclined to do." Aquinas's point is subtler: rational creatures share in eternal reason, and so their inclinations are normatively legible through practical reason. The law is rational before it is behavioral. This is why Aquinas can say that irrational creatures participate in eternal law too, but only rational creatures participate in a way properly called "law," because law belongs to reason (ST I-II q.91 a.2). If one ignores that rational participation, Thomistic natural law collapses into either biological description or divine command. Aquinas means neither. (New Advent)

1.2 Synderesis: first practical principles

Within that structure, synderesis names not a faculty but a natural habit of first practical principles. In ST I q.79 a.12 Aquinas is explicit: synderesis is "not a power but a habit." It contains the first practical principles bestowed by nature, "incites to good," and "murmurs at evil." These first practical principles are unerring at their own level; Aquinas says no one errs concerning them as first principles. This matters because synderesis is often misdescribed as a vague moral feeling or conscience. For Aquinas it is neither. It is closer to the inbuilt possession of the most universal practical truths by which later reasoning proceeds. (New Advent)

Aquinas immediately distinguishes synderesis from conscience. Conscience is not a power or a standing habit; it is an act—"knowledge applied to an individual case" (ST I q.79 a.13). Conscience witnesses, accuses, excuses, binds, or incites because it applies what is known to some particular deed here and now. That distinction becomes crucial for AI. Much of what current systems do when they self-critique, check a response against a policy, or apply a safety rule to a concrete prompt looks less like synderesis than like a thin analogue of conscience. Synderesis supplies universal starting points; conscience applies them. (New Advent)

1.3 Primary precepts and secondary precepts

The best-known Thomistic formula appears in ST I-II q.94 a.2: the first principle of practical reason is that good is to be done and pursued, and evil avoided. All other precepts of natural law are based on this. Aquinas then links natural-law content to ordered natural inclinations: self-preservation, procreation and care of offspring, and the distinctively rational inclinations to know truth about God and to live in society. These are not yet a complete ethics, but they are the intelligible roots from which more specific practical norms are derived. (New Advent)

Modern Thomistic summaries often speak of primary and secondary precepts. That shorthand is serviceable, but Aquinas's own language is more exact. He speaks of first principles, most general precepts, and then of "secondary and more detailed precepts" that are like proximate conclusions from those first principles (ST I-II q.94 a.6). He also insists that practical reason behaves differently from speculative reason: as one descends into matters of detail, defects, exceptions, and failures become more common. Goods held in trust should generally be returned, but not if returning them would arm an enemy against one's country. The more conditions one adds, the more ways a rule may fail in particulars. That is not skepticism about morality. It is Aquinas's account of why practical reasoning requires more than rule recitation. (New Advent)

Aquinas strengthens the point in ST I-II q.95 a.2, where he explains that more specific laws arise from natural law in two ways: as conclusions drawn from premises, or as determinations of general principles. "Do harm to no one" may yield "do not kill" as a conclusion. By contrast, that wrongdoers should be punished is more general, while the exact mode of punishment is a determination. This is one of the most useful Thomistic distinctions for AI design. Not every policy sits at the same level. Some are genuine derivations from general principles; others are contingent specifications chosen for governance, product, or operational reasons. (New Advent)

1.4 Prudence (prudentia) and the practical syllogism

If one stopped at synderesis and the precepts of natural law, one would have a theory of universal practical principles but not yet a theory of decision-making. Aquinas supplies that theory under prudence. Prudence is "right reason applied to action" (ST II-II q.47 a.8). It is not mere cleverness, prediction, or rule retrieval. Prudence requires knowledge of singulars, because action happens in singular matters. Aquinas says the prudent person must know both universal principles and the particular things to which they are applied; reason applies universal rules to particular cases, and without acquaintance with singulars no such application is possible (ST II-II q.47 a.3, a.8). (New Advent)

He further divides prudential activity into counsel, judgment, and command. Counsel is inquiry into what should be done; judgment assesses what inquiry has discovered; command applies that judgment to action, and Aquinas calls command the chief act of prudence because it stands nearest the end of practical reason. This is a hard-edged claim. Prudence is not just having good reasons in the abstract. It is reason that successfully gets all the way to action. (New Advent)

The practical syllogism belongs here. Aquinas says human action is directed by universal and particular knowledge; the practical syllogism concludes in a singular proposition or judgment about what is to be done (ST I-II q.76 a.1). He also notes that prudential reasoning proceeds from universal and singular premises (ST II-II q.49). But the Thomistic story does not end with a syllogistic conclusion. The conclusion of practical reason is followed by choice, and choice belongs substantially to the will, even though reason draws the practical conclusion (ST I-II q.13; q.14; q.8; I q.80). This is the pivot on which the AI analogy turns. A model may mimic the form of a practical syllogism. Whether it thereby performs a full practical act is another question. (New Advent)

2. Mapping Thomistic layers onto AI agent architectures

The most productive application of Thomism to AI is architectural, not anthropomorphic. Aquinas's account gives a way to describe layered norm-governed systems without pretending that a present-day model is a rational animal. If one treats an AI agent stack as a hierarchy of normative sources, derivations, and applications, the mapping becomes sharp enough to be useful. If one treats it as proof that models literally possess synderesis or prudence, the mapping collapses. (New Advent)

Thomistic layer	Aquinas's role	Closest AI analogue	What the analogy captures	Where it fails
Synderesis	Natural habit of first practical principles	System prompt / root instruction layer	Basal orientation before case reasoning	Synderesis is intrinsic and natural; prompts are authored and revisable
Primary / first precepts	Universal principles of practical reason	Constitution / model spec	High-level behavioral commitments	Constitutions are external governance artifacts, not participation in eternal law
Secondary precepts / determinations	More detailed conclusions and specifications	Product policies, developer rules, tool constraints	Domain-specific control and conflict resolution	They may reflect contingent design choices, not natural-law conclusions
Conscience	Knowledge applied to an individual case	Self-critique, policy check, case-specific review	Norm application to a concrete prompt	No inward witness; only text-level evaluation
Practical syllogism	Universal + singular premise yielding singular judgment	Visible reasoning trace or planner output	Structured derivation from norm to action	In Thomism, judgment must be followed by will-based choice
Prudence	Right reason applied to action, especially singulars	Best located in the wider sociotechnical system	Governance, oversight, escalation, command	Current models lack intellective appetite, embodied judgment, accountability

2.1 System prompts as synderesis—only in a thin sense

The user-supplied analogy—system prompts as synderesis—is intelligible, but it must be handled carefully. In contemporary deployed systems, high-authority instruction layers do play a basal orienting role. OpenAI describes the Model Spec as a formal public framework for model behavior, and its chain of command assigns higher authority to platform instructions, then developer instructions, then user instructions. Anthropic likewise publishes a constitution written primarily for the model as audience, and says that it directly shapes behavior and functions as final authority on intended values and conduct. At a purely functional level, these layers resemble a "first orientation" that shapes downstream reasoning before any particular task is considered. (OpenAI)

But in a strict Thomistic sense, system prompts are not synderesis. Synderesis is a natural habit of first practical principles, not a text string prepended to a context window. Aquinas places synderesis inside the rational creature as a naturally bestowed habit; natural law is promulgated because God has instilled it into the mind. A system prompt, by contrast, is an extrinsic, authored, revisable, and often product-specific instruction layer. So the stronger and more accurate claim is this: system prompts are synderesis-like in role, but closer in kind to promulgated law than to synderesis itself. They are better understood as an engineered analogue of externally notified practical directives than as an artificial natural habit. (New Advent)

2.2 Constitutional AI as primary precepts

The analogy becomes stronger with Constitutional AI. Anthropic's original paper defines Constitutional AI as a method where the only human oversight is a list of written rules or principles; the method uses supervised self-critique and revision, followed by reinforcement learning from AI feedback. The paper explicitly says those processes can leverage chain-of-thought-style reasoning for greater transparency and better judged performance. Anthropic's public constitution likewise presents a hierarchy of broad priorities—safe, ethical, compliant with guidelines, helpful—and then supplies more detailed guidance underneath those headings. This is structurally close to Aquinas's universal practical propositions: broad principles meant to govern a wide range of downstream judgments. (arXiv)

A further parallel appears in later work on constitutions. Kundu and colleagues ask whether a model can learn general ethical behavior from a single principle roughly equivalent to "do what's best for humanity." Their answer is qualified: large dialogue models can generalize from short constitutions, but more detailed constitutions improve fine-grained control over specific harms. That is strikingly Thomistic in form. Aquinas holds that the most general principles are stable, but more detailed practical rectitude becomes harder as one descends into particulars. A constitution-heavy alignment stack therefore looks like a technical attempt to recreate, in engineered form, Aquinas's distinction between very general principles and more specified practical guidance. (arXiv)

2.3 Contextual rules as secondary precepts and determinations

Once one leaves the constitutional layer, the analogy to secondary precepts becomes more direct. OpenAI's Model Spec explicitly introduces authority levels and concrete instructions for resolving conflicts across platform, developer, and user demands. Aquinas says human law develops from natural law partly as conclusions and partly as determinations. That is close to how modern agent stacks work. Some constraints feel like deductions from broad goals—e.g. do not facilitate serious harm if the broader system prioritizes harm prevention. Others are straightforward determinations—role permissions, tool-usage rules, product-specific limits, escalation protocols, and various governance conventions that could have been designed differently while still serving the same broad ends. (Model Spec)

This is one place where Thomistic vocabulary improves alignment discourse. "Policy" is too flat a word. A better question is: Is this rule a conclusion, or a determination? If it is a conclusion, one should be able to explain the derivation from a higher-order principle. If it is a determination, one should be able to say why this specification was chosen instead of another and under what authority. That distinction matters for audits, interpretability, and self-improving systems. It lets engineers separate genuine normative derivation from contingent governance choices. Aquinas's account of practical detail varying "in the majority of cases" but failing in some cases also warns against evaluating systems solely on slogan-level alignment with universal principles. (New Advent)

2.4 Model reasoning as practical syllogism

The analogy to the practical syllogism is the most visible and the most tempting. A model given a constitution and a case often produces an answer with the following structure: major premise — a general principle or policy, minor premise — a description of the present case, conclusion — therefore this response or action is appropriate.

That is recognizably Thomistic in form. Aquinas says practical reason works from universal and particular knowledge, and that the conclusion of the practical syllogism is a singular proposition concerning action. Constitutional AI's self-critique and revision procedures intensify this resemblance because the model is literally asked to examine a candidate answer against higher-level rules and produce a corrected one. (New Advent)

Still, one must be precise about which Thomistic category is being mirrored. A model's case-specific self-critique is usually closer to conscience—knowledge applied to an individual case—than to synderesis. Synderesis supplies the universal first principles. Conscience applies them and judges this act here and now. Practical-syllogistic outputs sit at that application level. They are the most plausible place to say that an AI system "reasons" in a thin practical sense. But they are not yet prudence in the full Thomistic sense, because prudence includes command and presupposes a will that can choose means toward an apprehended good. (New Advent)

3. Where the analogy breaks down

3.1 Natural teleology versus extrinsic steering

The first break is metaphysical, not cosmetic. Aquinas's natural law presupposes that rational creatures are internally ordered toward their proper acts and ends by participation in eternal law. By contrast, contemporary AI systems are steered by external governance artifacts: system messages, constitutions, reward shaping, model specs, and deployment constraints. OpenAI says this with unusual clarity: benefiting humanity is OpenAI's goal, not a goal the model is meant to pursue autonomously; the model is meant instead to follow a chain of command. Tabaczek, writing from an Aristotelian-Thomistic perspective, makes a similar point in different language: AI goal-directedness is better understood as externally programmed teleonomy than as the intrinsic teleology proper to natural agents. (New Advent)

This difference is decisive. In Aquinas, the first practical principles belong to the agent as rational participant in divine order. In current AI systems, the highest-order norms are imposed by builders and deployers. That does not make them unimportant; on the contrary, it makes them central. But it means that AI alignment more closely resembles lawgiving and governance than natural-law participation. The practical relevance is immediate: norms in AI systems are fragile in a way synderesis is not. They can be rewritten, truncated, overridden, or inconsistently layered. Thomism helps diagnose that fragility precisely because its original categories are thicker than the engineering analogues. (New Advent)

3.2 The absence of intellective appetite

The second break is Aquinas's doctrine of the will. For him, the will is a rational appetite. The intellectual appetite tends toward individual things under the universal aspect of the good; choice follows reason's judgment but belongs substantially to the appetitive power, not to reason alone. This is why practical reasoning in Aquinas is not exhausted by valid inferential form. A completed practical act includes appetitive movement toward the chosen means. Without intellective appetite, there is no full-blooded Thomistic choice. (New Advent)

Current language models do not satisfy that condition. They can represent goods linguistically; they do not literally appetite them in Aquinas's sense. Their outputs are not acts of a rational appetite tending toward an apprehended good; they are generated continuations constrained by learned distributions and runtime instructions. This is why claims that LLMs possess prudence or moral virtue in the classical sense should be rejected. Noller argues that constitutionally trained models can exhibit stable, norm-guided behavior, but that this does not amount to virtue or moral character in the Aristotelian sense, because virtue remains tied to embodied agency, affectivity, and practical judgment. Tabaczek reaches an even sharper Thomistic conclusion: properly human virtues are highly unlikely to be attributable to AI, though one might speak more cautiously of "machine-based" analogues in weak AI. (Springer Link)

The same point blocks a common shortcut in AI ethics. It is not enough to say that a model "acts as if" it values honesty, justice, or safety. For Aquinas, virtuous action is not defined only by externally right behavior. It also depends on the kind of subject that acts, the powers that are perfected, and the way the agent is disposed toward the good. Thomistic language can still be useful, but mainly as analogy or instrumental description, not literal attribution. (New Advent)

3.3 Moral luck in stochastic systems

The third break is what the prompt rightly calls the problem of moral luck in stochastic systems. Current LLMs produce next-token probability distributions and then generate by iteratively selecting from those distributions. Decoding settings and penalties can directly modify logits and therefore alter which continuations are likely. In practice, this means that whether a norm is expressed clearly, weakly, or not at all can depend on contingent features of sampling, context-window composition, or other runtime details. By analogy to moral luck, the outward moral profile of a particular run can be affected by contingencies that are not reducible to a stable, self-possessed character in the Thomistic sense. (Hugging Face)

This does not imply that model behavior is random in a colloquial sense. It does imply that the unit of responsibility is hard to locate if one tries to use classical virtue language. A virtue is a stable disposition perfecting a power of an agent. Yet model outputs are realized through a pipeline involving pretrained weights, alignment tuning, system instructions, retrieval context, decoding rules, tool calls, and possibly other agents supervising the process. The "character" of the system is distributed across the sociotechnical stack. That is one reason Noller reframes alignment as an ongoing practice of mediated responsibility rather than a property of the artificial agent itself. (Forum Philosophicum)

The engineering consequence is stark. If you evaluate a model only at the level of one sampled answer, you can mistake output variance for norm instability or, conversely, mistake a lucky answer for reliable alignment. A Thomistic lens encourages a different emphasis: examine whether the system can reliably derive lower-level judgments from higher-order commitments across varying singular cases. That is closer to testing the analogues of habit, conscience, and prudential application than merely testing for isolated correct utterances. (New Advent)

3.4 Can probabilistic outputs constitute genuine practical reasoning?

This question needs a split answer.

In a thin functional sense, yes: probabilistic systems can produce outputs that instantiate the structure of practical reasoning. A model can cite a norm, identify salient particulars, and derive a response that is intelligibly connected to both. Constitutional AI explicitly trains systems to critique outputs in light of written principles, and Aquinas's account of the practical syllogism is flexible enough that one can see a formal resemblance here. If all one means by practical reasoning is publicly inspectable norm-sensitive inference about what should be done, present systems can sometimes achieve it. (arXiv)

In a strict Thomistic sense, no: probabilistic outputs do not by themselves constitute genuine practical reasoning, because Aquinas's practical reason culminates in judgment ordered to choice, and choice belongs to will. A string of tokens that mirrors a practical syllogism is not yet a practical act in the full sense unless it belongs to a subject with intellective appetite and the capacity for command. For Aquinas, practical reason is inseparable from the agent's teleological and appetitive structure. Current models lack that structure. Therefore their "reasoning" is, at best, a simulation or instrumental surrogate of practical reasoning, not its literal realization. (New Advent)

There is, however, a weaker claim worth defending: systems may still function as artificial moral assistants even if they are not moral agents. Recent work on artificial moral assistants states the point directly: to produce valid moral advice, it is not necessary to be an ethical agent, because the validity of ethical reasoning is independent of its origin. That distinction is philosophically important and engineering-relevant. One can reject the attribution of prudence or virtue to the model while still taking seriously its capacity to support human deliberation, flag conflicts, surface considerations, and generate ethically relevant arguments. (arXiv)

This distinction also helps with current alignment critiques. Schuster and Kilov argue that crowdsourcing, RLHF, and constitutional AI do not adequately accommodate reasonable moral disagreement in controversial decisions. Millière argues that current alignment methods often yield shallow alignment—behavioral regularities without deep normative resolution. Those critiques fit neatly with a Thomistic diagnosis: present systems can often reproduce the surface form of moral reasoning without possessing the integrated practical wisdom, appetitive ordering, and command that would make the reasoning fully their own. (Springer Link)

4. Existing academic work on virtue ethics in AI alignment

The literature here is real but uneven. There is a modest body of work on virtue ethics and AI, a growing body on constitutional or norm-guided training, and only a comparatively thin layer that is explicitly Aristotelian-Thomistic. There is not, as of now, a mature scholarly tradition that directly maps synderesis → system prompt, secondary precepts → policy layer, and prudence → agent planner. That specific mapping is mostly a constructive synthesis of Aquinas with contemporary alignment practice, not an established school of interpretation. (arXiv)

Work	Main contribution	Why it matters here
Berberich & Diepold, "The Virtuous Machine" (2018) (arXiv)	Argues Aristotelian virtue ethics fits modern AI because of learning from experience; proposes imitation learning from moral exemplars and suggests virtues like temperance/friendship could mitigate control problems.	Early attempt to treat alignment as cultivation of dispositions rather than compliance with flat rules.
Govindarajulu, Bringsjord, Ghosh, "Toward the Engineering of Virtuous Machines" (2018) (arXiv)	Formalizes exemplar-based virtue learning for machine ethics.	Shows one concrete way virtue ethics can be operationalized computationally.
Hagendorff, "A Virtue-Based Framework…" (2022) (Springer Link)	Treats virtue ethics as a complement to principle-based AI ethics, with justice, honesty, responsibility, care, prudence, and fortitude cultivated especially at the organizational level.	Important corrective: prudence may belong more to AI institutions than to models.
Tabaczek, "Virtuous AI?" (2024) (Forum Philosophicum)	Gives an explicitly Aristotelian-Thomistic treatment; argues properly human virtues are not attributable to AI, though "machine-based" analogues may be discussable for weak AI.	Closest direct companion to this entry.
Noller, "Artificial moral characters" (2026) (Springer Link)	Reads Constitutional AI through virtue ethics and 4E cognition; stable norm-guided patterns are not virtue or character in the Aristotelian sense.	Strongest recent argument against naïve character attribution to LLMs.
Graves, "AI practical wisdom and compassion" (2026) (Springer Link)	Proposes practical wisdom grounded in compassion as an alignment orientation, especially in healthcare.	Illustrates the most ambitious contemporary attempt to operationalize a virtue-like alignment target.

The early virtue-ethics work is striking for what it gets right and what it leaves underdeveloped. Berberich and Diepold correctly saw that virtue language is attractive for machine ethics because modern AI learns from experience rather than simply executing static symbolic rules. Govindarajulu and collaborators pushed in a more formal direction by modeling exemplar-based learning. What both approaches show is that virtue ethics is not inherently "too fuzzy" for technical work. What they do not show is that a model trained in these ways would thereby possess virtue in the full classical sense. Thomistic concerns about appetite, teleology, embodiment, and the subject of habit remain. (arXiv)

Hagendorff's intervention is useful precisely because it relocates virtue. Rather than pretending that AI systems themselves straightforwardly bear virtues, he argues that virtue ethics can complement principle-based AI ethics by shaping the motivational and institutional culture of organizations building AI. For a Thomistic reader, this is attractive. Prudence, justice, honesty, and care may be most plausibly attributed to the human community designing, deploying, and auditing the system. That is much closer to Aquinas's understanding of virtue as perfection of powers in a subject than the looser habit of attributing "character" to a model because its outputs appear stable. (Springer Link)

Tabaczek is the most directly relevant contemporary source for a Thomistic framing. His analysis is blunt: from within Thomistic ontology, specifically human virtues are not genuinely reproducible in AI, because virtue belongs properly to a subject with the relevant powers and intrinsic teleology. He is open to speaking of "machine-based" prudence or other analogues in weak AI, but only if the terminology is carefully qualified and kept distinct from properly human virtues. That caution is exactly the right one. It allows architectural and functional comparisons without quietly smuggling in claims about ontology or moral status that the Thomistic framework itself would reject. (Forum Philosophicum)

Noller's recent paper is important because it addresses constitutional AI directly. His claim is not that constitutionally trained models are normatively irrelevant. It is that stable norm-guided behavior does not yet constitute virtue or moral character in the Aristotelian sense. He therefore shifts analysis away from "Is Claude virtuous?" toward "How are human moral commitments externalized and stabilized in technical infrastructures?" That move is deeply compatible with the argument of this entry. It treats constitutional models as loci where human normative order is encoded, mediated, and enforced—not as straightforward bearers of classical virtue. (Springer Link)

The more ambitious side of the literature appears in work like Graves's discussion of artificial practical wisdom and compassion. Such work is valuable because it tests whether a virtue-ethical vocabulary can do real design work in high-stakes domains. The weak point is that "practical wisdom" is often used there in a thinner sense than in Aquinas: closer to ethically informed context sensitivity and deliberative competence than to the full Thomistic virtue of prudence. That thinner use may be defensible for engineering, but it should be named as such. Otherwise "practical wisdom" becomes a flattering label for sophisticated policy following. (Springer Link)

Finally, the critical literature on moral disagreement and shallow alignment underscores why virtue language became tempting in the first place. If principle lists and preference datasets cannot fully resolve controversial cases, then a richer account of judgment seems necessary. But the existence of that need does not prove that current models have acquired prudence. It proves only that alignment based on rule enumeration and preference fitting runs into the classic problem Aquinas already saw: general norms are not self-applying in singular, contingent, high-stakes situations. (Springer Link)

5. What AI engineers can take from a Thomistic lens

5.1 Separate universal commitments from determinations

A Thomistic approach recommends an explicit norm stack. Do not collapse everything into "the policy." Separate the constitutional layer from lower-level determinations. Ask which rules are meant to function like universal practical commitments and which are contingent specifications chosen for governance. Contemporary constitutional-AI work already suggests that broad principles and detailed rules do different kinds of work; Aquinas gives the conceptual vocabulary for that difference. Systems that fail to separate them become hard to audit and harder to revise coherently. (arXiv)

5.2 Evaluate application to singulars, not just recitation of principles

Aquinas insists that prudence is about singulars and that practical detail is where failures multiply. That implies an evaluation lesson: test whether an agent can apply high-level norms across varied particulars, edge cases, and conflicting conditions. A model that can quote its constitution is not thereby well aligned. What matters is whether it can move from universal premise to apt judgment in cases where details matter and where simple policy retrieval is insufficient. This is one reason red-teaming, adversarial evaluation, and scenario diversity are not optional extras. They are the nearest current analogue to testing the prudential application of norms. (New Advent)

5.3 Locate prudence in the sociotechnical system

The chief act of prudence is command. In current AI deployments, the point of command usually belongs not to the base model but to the surrounding stack: orchestration logic, tool permissions, humans with override authority, institutional review processes, and deployment governance. A Thomistic frame therefore pushes against the habit of treating the model alone as the moral subject. The more plausible bearer of prudence is the human–technical ensemble that sets ends, authorizes means, handles exceptions, and accepts responsibility. Noller's "mediated responsibility" language is a good contemporary expression of this point. (New Advent)

5.4 Treat self-improvement as critique under law, not self-legislation

Constitutional AI is attractive because it uses AI to critique AI under a written constitution. Thomistically, that is sensible so long as higher-order principles remain distinguishable from the lower-order reasoning they govern. But one should resist letting self-critique blur into self-legislation. Aquinas's distinction between first principles, derived conclusions, and determinations suggests a design discipline: higher-order normative commitments must remain externally inspectable and govern the revision loop rather than being silently rewritten by it. For self-improving systems, this is not a small point. It is the difference between improvement under law and uncontrolled drift in the source of law. (arXiv)

5.5 Prefer "moral assistant" to "moral agent"

For present systems, the strongest defensible claim is usually not "this agent is prudent" or "this model has virtue," but "this system can assist moral reasoning under supervised constraints." That vocabulary is better in both directions. It avoids anthropomorphic inflation, and it does not undersell what the system can do. A good artificial moral assistant can still surface considerations, explain tensions, help maintain consistency, and improve human deliberation. Thomistically, that role is easier to defend because it does not require us to pretend that probabilistic output generation is identical with intellective appetite and practical choice. (arXiv)

Conclusion

Applied to AI agent decision-making, Thomistic natural law theory is best understood as a grammar of alignment layers. Synderesis highlights the importance of first practical principles. The theory of natural-law precepts clarifies the difference between universal norms and more detailed derived rules. Conscience explains case-specific application. Prudence explains why singular judgment and command matter more than abstract norm recital. And the practical syllogism explains why reasoned action is not just rule storage but movement from universal principle to concrete judgment. On those points, Aquinas is unexpectedly useful for AI engineering. (New Advent)

But the same framework also tells us exactly where to stop. Present AI systems do not participate in eternal law, do not possess a natural habit of first principles, do not have intellective appetite, and do not choose means under a will oriented to the good. They can be governed, constrained, and sometimes made to reason as if they were practical agents. They can perhaps become powerful moral assistants. They do not thereby become Thomistic moral subjects. Any account that forgets that boundary will misdescribe both Aquinas and the systems it hopes to illuminate. (New Advent)

Natural companion entries

Synderesis Natural Law Theory Prudence (Prudentia) Practical Reason and the Practical Syllogism Constitutional AI Model Spec and Instruction Hierarchies Artificial Moral Agency Moral Luck and Stochastic Systems Virtue Ethics in AI Alignment AI Value Alignment

Selected primary texts and contemporary sources

Primary Thomistic loci: Summa Theologiae I q.79 a.12–13; I-II q.90 a.1–4; I-II q.91 a.2; I-II q.94 a.2, a.4–6; I-II q.95 a.2; I-II q.13–14; I-II q.76 a.1; II-II q.47 a.3, a.8; II-II q.49. (New Advent)

Key contemporary sources for the AI side: Anthropic's Constitutional AI paper; Anthropic's public Claude's Constitution; OpenAI's Model Spec materials; Kundu et al. on specific versus general constitutions; Berberich & Diepold on the virtuous machine; Govindarajulu et al. on engineering virtuous machines; Hagendorff on virtue-based AI ethics; Tabaczek on virtuous AI from an Aristotelian-Thomistic perspective; Noller on artificial moral character; Graves on AI practical wisdom and compassion; Schuster and Kilov on moral disagreement; Millière on shallow alignment; and recent work on artificial moral assistants. (arXiv)