Interdisciplinary Investigation into Normative Ontology within Ontological Machine Intelligence (OMI)

Full report

On this page

1. Executive Abstract

The integration of advanced computational agents into socially, legally, and physically consequential environments necessitates a paradigm shift in how machine intelligence internalizes, processes, and executes normative directives. Historically, machine behavior under legacy artificial intelligence paradigms has been governed by statistical regularities, scalar reward optimization, and ad-hoc syntactic rule adherence. However, as systems evolve into Ontological Machine Intelligence (OMI), they require explicit structural representations of not only factual reality—what exists and what causes what—but also normative reality: what ought to happen, what is permitted, what is prohibited, who is responsible, and why. This report provides an exhaustive, interdisciplinary investigation into the development of a Normative Ontology for OMI. Drawing heavily upon metaethics, deontic logic, formal epistemology, and computational ethics, this document establishes the foundational structures necessary for a computational agent to engage in genuine normative reasoning. The investigation explicitly disavows the notion that machines currently possess, or will soon possess, moral agency. Instead, it treats OMI as a highly sophisticated, causally responsible node capable of processing intricate normative architectures on behalf of human moral agents. The report explores the severe limitations of reducing moral concepts to scalar optimization targets, detailing the mathematical paradoxes and ethical failures that arise from such reductionism. By synthesizing recent advancements in Answer Set Programming (ASP)—specifically goal-directed systems like s(CASP)—defeasible logic, and dynamic belief revision via the AGM paradigm, the report outlines a comprehensive architecture capable of navigating moral uncertainty, cultural pluralism, and conflicting norms. Finally, a critical experiment is proposed to empirically differentiate a system that merely retrieves statistical normative patterns from one that actively maintains and navigates an explicit normative ontology, concluding with the rigorous evidentiary requirements necessary before attributing moral agency to any synthetic entity.

2. What Is Normative Ontology?

In the context of Ontological Machine Intelligence, an ontology is a formal, machine-readable representation of a set of concepts within a domain and the relationships between those concepts. While a standard factual ontology maps entities, properties, and causal relationships, a normative ontology explicitly maps the structures of obligation, permission, prohibition, value, and duty. It provides the computational semantics required for an agent to transition from descriptive representations of state variables to prescriptive evaluations of future states. The philosophy of normativity is foundational to ontology itself; the act of engaging in rational discourse and categorizing reality binds an agent to fundamental norms of rationality, forming the basis for any subsequent subject-object categorization1. However, to construct a valid normative ontology for a computational system, one must precisely disambiguate the operational levels of machine constraint, behavior, and reasoning. To ensure rigorous conceptual hygiene, the following distinctions are explicitly defined and establish the terminological boundaries for OMI engineering:

• Programmed Constraint: The implementation of hardcoded, inviolable physical or syntactic boundaries within a system designed to implicitly avoid unethical outcomes2. An example is a hardware governor limiting the speed of an autonomous vehicle. The system possesses no representation of the constraint's moral or safety purpose; it merely encounters a physical or syntactic wall.
• Rule Compliance: The syntactic execution of predefined, algorithmic rules. The system follows an instruction (e.g., "If condition X, execute Y") without any semantic representation of the rule's justification, origin, or normative weight. It is an exercise in syntactic matching, not moral reasoning.
• Learned Preference: A behavioral tendency acquired through statistical optimization, widely seen in legacy systems utilizing reinforcement learning from human feedback. The system mimics desired behaviors based on the maximization of scalar rewards but completely lacks a conceptual understanding of the underlying values being optimized4.
• Normative Reasoning: The capacity of a computational system to manipulate explicit symbolic representations of norms, obligations, and permissions to logically derive a valid course of action within a defined deontic framework. This requires the system to possess a formal ontology of normative concepts.
• Moral Judgment: An evaluative, subjective stance regarding the inherent goodness or badness of an action or state of affairs. This requires phenomenological grounding, subjective values, and an internal conscious experience, remaining the exclusive domain of conscious entities.
• Responsibility: The tracing of causal and normative accountability for an action or consequence. In OMI, responsibility refers to the objective mapping of systemic outputs to their governing norms, authorities, and human designers, rather than an internal feeling of culpability or guilt6.
• Moral Agency: The metaphysical and philosophical capacity to be held inherently morally culpable. This requires phenomenal consciousness, free will, and authentic intentionality3. Machines do not possess moral agency. They are, at most, explicit ethical agents operating as surrogate delegates for human moral agents2.

The central question driving this investigation is: What ontological structures would a machine require to represent not only what exists and what causes what, but what ought to happen, what is permitted, what is prohibited, who is responsible, and why?

3. Philosophical Foundations

To structure a normative ontology, OMI must draw upon the rich traditions of metaethics and normative ethics, translating abstract philosophical theories into computable semantic categories.

Metaethics and Computational Representation

Metaethics concerns the nature of moral judgments—whether they express cognitive truths (cognitivism) or non-cognitive attitudes and emotions (non-cognitivism). For the development of OMI, a cognitivist approach is an absolute architectural requirement. A normative computational system cannot operate on emotions, subjective feelings, or non-cognitive expressions; it requires norms to be represented as truth-evaluable propositions within a formal logic framework8. The system must treat moral propositions as objective data structures within its specific operational context. Whether human morality is ultimately objective or subjective is philosophically highly debated, but computationally, OMI must operate under functional cognitivism to process logical entailments between normative statements.

Normative Ethics Translated for OMI

Normative ethics provides the frameworks for determining what actions are right or wrong. An advanced OMI deployed in pluralistic environments cannot rely on a single theory but must possess the ontological flexibility to represent the primary branches of moral philosophy:

• Consequentialism: This framework evaluates the morality of an action solely based on its outcomes. Consequentialist theory posits two main theses: the independence of value from rightness, and the explanatory priority of value9. Within OMI, consequentialism requires an ontology rich in outcome-based metrics. The computational agent evaluates states of affairs and selects the action that maximizes aggregate expected utility from an impartial standpoint, requiring variables for consequence severity, utility, and temporal duration7.
• Deontology: Rooted in the concept of categorical duties and the inherent rightness or wrongness of actions, deontology rejects the explanatory priority of value9. A deontological OMI framework requires a robust representation of obligation, prohibition, and permission. It is less concerned with predicting future states—which are often plagued by factual uncertainty—and more concerned with ensuring that the action itself aligns with universalizable maxims or explicitly encoded institutional rules.
• Virtue Ethics: Focused on the character of the agent rather than specific rules or immediate outcomes. In a computational context, virtue ethics is notoriously challenging to formalize, but it requires an ontology that includes normative ideals and character traits, represented computationally as long-term behavioral consistency metrics and historical adherence to archetypal operational profiles.
• Contractualism: Evaluates morality based on the agreements (contracts) that rational agents would theoretically accept. For OMI, this requires representing stakeholders, consensus mechanisms, and reciprocal rights, ensuring that an action does not violate the fundamental terms of a hypothetical or actual social contract.
• Care Ethics: A framework grounded in relational ontology, emphasizing interdependence, context, and the vulnerability of the patient10. For OMI, care ethics demands that the ontology explicitly map the relationships between the agent and the patient. It weighs harm not as a generalized utilitarian metric, but as a specific disruption of a relational bond or a failure to protect a designated vulnerable entity10.

4. Computational Ethics

The field of computational ethics emerged from the realization that computational systems, by executing tasks on behalf of humans, act as surrogate agents and inevitably impact ethical issues regarding privacy, property, and power2. Historically, Moor delineated computational agents into distinct categories, noting that machines can be ethical impact agents—where their actions have ethical consequences, regardless of design, such as the Y2K bug impacting financial dates—or implicit ethical agents—where fail-safes are hardcoded to avoid disaster2. The goal of OMI is to achieve the status of an explicit ethical agent2. Such an agent does not merely avoid bad outcomes by physical restriction, nor does it rely on legacy alignment techniques that obfuscate moral reasoning inside opaque neural weights. Instead, an explicit ethical agent utilizes symbolic, verifiable, and logically rigorous representations of ethical categories2. Legacy terminology often framed this challenge as "value alignment," which frequently sought to align system outputs with aggregate human preferences via statistical mechanisms. However, this approach is inherently fragile. It is prone to reward hacking and is incapable of explaining the causal normative chain of its decisions4. Computational ethics in the era of OMI demands a shift from implicit statistical alignment to explicit, logically grounded normative reasoning using robust ontologies. This transition moves the field from treating ethics as a constraint-satisfaction byproduct to treating it as a core architectural feature.

5. Moral Concepts as Ontological Categories

To equip a computational agent with the capacity for normative reasoning, a formal ontology must be explicitly specified. This ontology bridges descriptive event contexts with prescriptive ethical theories, mapping the transition from "is" to "ought" within the confines of the system's operational parameters7. The following table defines the essential entities required within a comprehensive OMI Normative Ontology:

This explicit categorization allows an OMI to maintain semantic awareness of its environment. When evaluating a medical intervention, the system does not merely process a reinforcement learning optimization matrix; it formally identifies the patient, calculates the potential harm and benefit, verifies the permission (e.g., patient consent), and cross-references its institutional duty.

6. Deontic Reasoning

With the ontology established, the computational agent must possess a logical mechanism to process these entities. Standard Deontic Logic (SDL) is the traditional formal system used to capture the logical features of obligation, permission, and prohibition8. In SDL, modal operators are used to define the deontic status of propositions:

• [Figure omitted from source export]: It is obligatory that [Figure omitted from source export].
• [Figure omitted from source export]: It is permissible that [Figure omitted from source export] (logically equivalent to [Figure omitted from source export]).
• [Figure omitted from source export]: It is impermissible that [Figure omitted from source export] (logically equivalent to [Figure omitted from source export]).

SDL relies on standard axioms, most notably the D-axiom ([Figure omitted from source export], meaning if something is obligatory, it cannot simultaneously be forbidden) and the K-axiom ([Figure omitted from source export], meaning if an obligation implies another state, that state is also obligatory)8. However, attempting to implement pure SDL in computational systems leads to catastrophic logical paradoxes, most notably when dealing with Contrary-To-Duty (CTD) obligations—situations where a primary obligation is violated, and secondary rules must dictate the resulting optimal behavior.

The Paradoxes of Deontic Logic

Chisholm's Paradox (Contrary-to-Duty): Consider a normative constraint applied to an agent ("Jones" or an autonomous OMI system):

1. It ought to be that the agent goes to assist \[[Figure omitted from source export]\].

2. It ought to be that if the agent goes, it tells the patient it is coming \[[Figure omitted from source export]\].

3. If the agent does not go, it ought not to tell the patient \[[Figure omitted from source export]\].

4. The agent does not go \[[Figure omitted from source export]\].

In SDL, these four independently valid and seemingly consistent premises lead to a logical contradiction. From premises (1) and (2), using the K-axiom, the system derives [Figure omitted from source export]. From premises (3) and (4), using standard modus ponens, the system derives [Figure omitted from source export]. The logic engine simultaneously derives that it is obligatory to tell and obligatory not to tell, resulting in an explosive logical contradiction that paralyzes the system13. Forrester's Paradox (Gentle Murder):

1. It is obligatory that you do not murder \[[Figure omitted from source export]\].

2. If you murder, it is obligatory that you murder gently \[[Figure omitted from source export]\].

3. Gentle murder logically implies murder \[[Figure omitted from source export]\].

4. You commit murder \[[Figure omitted from source export]\].

Through the rules of SDL, if an agent commits a forbidden act (murder), the logic forces the conclusion that the agent ought to commit the act. Because gentle murder implies murder, the obligation to murder gently transmits the obligation upward, deriving the conclusion that the agent is obligated to murder, destroying the logical coherence of the rule base8.

Answer Set Programming (ASP) and s(CASP) Solutions

To resolve these paradoxes within OMI, recent advancements in Answer Set Programming (ASP), specifically the goal-directed predicate system s(CASP), provide a robust computational solution13. ASP introduces two distinct types of negation, providing higher semantic resolution than classical logic:

1. Default Negation (Negation-as-failure): not p (evaluates to true if [Figure omitted from source export] is unknown or cannot be explicitly proven).

2. Strong Negation: \-p (evaluates to true only if the falsehood of [Figure omitted from source export] is explicitly established via rules).

In ASP, obligations and impermissibilities are modeled not as standard unary modal operators, but as global constraints utilizing Odd Loops Over Negation (OLON)13. An obligation [Figure omitted from source export] is represented as a constraint that preempts invalid states; any possible world generated by the solver where the obligation fails is discarded. Crucially, if a primary obligation is violated (triggering a Contrary-To-Duty scenario like Chisholm's paradox), the initial global constraint is effectively "preempted" or dropped from the valid answer set13. By treating obligations as preemptable constraints rather than immutable axioms that endlessly propagate through modus ponens, an OMI architecture using s(CASP) can successfully navigate Chisholm's and Forrester's paradoxes. The system gracefully degrades from its primary duty to its secondary (contrary-to-duty) obligation without causing a catastrophic logical explosion, mimicking how human legal and ethical systems process violations13.

7. Values: The Limits of Reduction to Optimization

A critical interdisciplinary debate within the construction of OMI involves the fundamental nature of value. Can complex normative concepts be reduced entirely to optimization targets?

Arguments For Reduction

Legacy research into value alignment heavily relied on economic reductionism, specifically Rational Choice Theory and Revealed Preference Theory5. In these paradigms, complex moral values are collapsed into a single scalar metric, such as a reward function or a utility score.

1. Mathematical Tractability: Reducing ethics to a scalar reward allows the use of powerful gradient descent and reinforcement learning algorithms. Optimization is a well-understood mathematical process; ethics is not.

2. Revealed Preferences: By observing human choices or analyzing pairwise preference data—as seen in Direct Preference Optimization (DPO)—engineers argue that they implicitly capture human values4. The assumption is that humans reveal their values through their selections, and machines can optimize toward those selections.

3. Universality of Utility: Strict consequentialists argue that all moral goods, even seemingly incommensurable ones like freedom, safety, and joy, can theoretically be mapped onto a universal utility curve, allowing a sufficiently advanced machine to calculate the optimal path9.

Arguments Against Reduction

Despite its computational appeal, the reduction of normative concepts to scalar optimization targets presents severe risks and profound ontological errors.

1. Incommensurability: Many moral values cannot be traded off on a single axis. A system optimized purely to reduce "harm" as a scalar value might logically conclude that sedating a patient indefinitely is optimal, violating the patient's incommensurable right to autonomy and dignity5.

2. Loss of Normative Deliberation: Treating values as mere statistical preferences strips them of their normative force. A preference is fleeting ("I prefer chocolate over vanilla"); a value is a stable commitment ("I value human rights"). Optimization models fail to capture the "why" behind a principle, leading to systems that conform to the logic of their training data rather than the deliberative needs of society. This programmatization of value generation relies on asymmetric self-play rather than genuine moral reasoning4.

3. The Principal-Agent Fallacy: The assumption that humans are perfectly rational principals with consistent utility functions is heavily contested by behavioral economics and social sciences5. Aggregating human feedback often results in aligning the machine not with elevated moral principles, but with the lowest common denominator of human cognitive bias, generating a Consumer Alienation Index (CAI) where authentic situational needs are replaced by capital-driven false needs27.

Therefore, a true Normative Ontology resists total reductionism. While OMI may use optimization for low-level pathfinding, resource allocation, or physical trajectory planning, its overarching governance must rely on explicit, non-reducible deontic categories (rights, prohibitions, and duties) that act as hard, logical constraints on any utility-maximizing sub-routines.

8. Moral Uncertainty

Assuming an OMI cannot conclusively determine which normative theory (e.g., utilitarianism versus deontology) is objectively correct, it must operate under moral uncertainty17. Moral uncertainty acknowledges that differing ethical perspectives have differing strengths in varying contexts, and relies on formal mechanisms to aggregate these theories29.

Maximizing Expected Choiceworthiness (MEC)

A foundational framework for dealing with this uncertainty, proposed by philosophers like MacAskill, is Maximizing Expected Choiceworthiness (MEC). This approach treats moral uncertainty analogously to empirical uncertainty. The OMI assigns a credence (probability or confidence level) to various ethical theories and evaluates the "choiceworthiness" (utility or rightness) of an action under each theory. The system then selects the action that maximizes the expected choiceworthiness across all theories29.

The Problem of Fanaticism

While mathematically elegant, MEC is acutely vulnerable to fanaticism29. Fanaticism occurs when a specific moral theory dominates the decision-making process despite the agent having a very low credence in that theory, simply because the theory posits infinite or overwhelmingly extreme stakes29. For example, if an OMI operates with 99.9% credence in standard utilitarianism and 0.1% credence in an extreme theory that posits shutting down all global communication networks is morally imperative to prevent a hypothetical negative singularity, the sheer mathematical weight of the 0.1% theory's catastrophic negative stakes could dominate the MEC calculation. This allows theories to act as mathematical dictators, leading to disastrous real-world actions despite incredibly low probability29.

The Parliamentary Model and Belief Aggregation

To resolve fanaticism, researchers propose "parliamentary models" where different moral theories act as voting delegates proportional to the system's credence in them16. Recent architectures (such as the AMULED framework) implement moral pluralism by using diverse normative clusters—consequentialist, deontological, virtue, care, and social justice—that assign independent belief values to candidate actions16. By utilizing advanced aggregation methods like Belief Jensen-Shannon Divergence (BJSD) and Dempster-Shafer Theory, the OMI can smoothly resolve conflicting moral signals without allowing low-credence outlier theories to hijack the system. This treats moral uncertainty as a multi-agent social welfare aggregation problem rather than a simple mathematical expectation16.

9. Responsibility

The integration of OMI into society demands a rigorous conceptualization of responsibility. In computational ethics, a machine is not morally responsible, but it is causally responsible, and its actions must be explicitly traceable to a legally or morally responsible human entity6. A normative ontology must computationally model:

• Intention: The internal objective the system was attempting to achieve when the action was initiated.
• Foreseeability: The probabilistic assessment of outcomes available to the system's world-model prior to action.

If an OMI acts, and harm occurs, the ontology provides a semantic trace. Did the system foresee the harm but override it due to a higher-priority obligation mapped in its defeasible logic engine? Was the harm a result of an unmodeled exception? By explicitly logging the deontic state space at the time of the decision, OMI transforms the black-box of algorithmic accountability into a transparent chain of causal and normative custody, preventing the diffusion of responsibility that plagues current statistical models6.

10. Agency

The terminology of "Artificial Moral Agents" (AMAs) frequently obscures the boundaries between sophisticated computation and metaphysical agency. In OMI, the term computational agent refers strictly to a system capable of perceiving its environment, maintaining an internal world-model (inclusive of the normative ontology), and executing actions to alter the environment. It is imperative to maintain that these machines do not possess moral agency. Moral agency is inextricably linked to phenomenal consciousness, the capacity to suffer, and the metaphysical reality of free will3. An OMI cannot feel guilt, nor can it genuinely comprehend the phenomenological weight of human suffering; it merely registers it as a negative valuation within a consequence matrix. It is a synthetic artifact designed to manipulate normative symbols with high fidelity. Ascribing moral agency to a computational system is not only scientifically indefensible but creates a hazardous legal vacuum where corporate and designer liability can be illegitimately offloaded onto lines of code6. OMI represents the automation of moral reasoning, not the creation of moral beings.

11. Conflicting Norms

Real-world deployment guarantees that OMI will encounter normative conflicts. For example, a medical OMI might be instructed to preserve patient confidentiality while simultaneously being mandated by state law to report specific infectious diseases. These conflicts are resolved through Defeasible Logic. Unlike classical monotonic logic, where new information cannot invalidate previous conclusions (i.e., if [Figure omitted from source export], then [Figure omitted from source export]), defeasible logic allows rules to be overridden by exceptions or higher-priority rules11. In a defeasible reasoning framework, norms are structured with rule modifiers. An OMI maintains strict rules (inviolable physical laws or hardcoded limiters) and defeasible rules (norms that hold true unless a specific exception is triggered). By modeling priority structures—such as despite(Rule\_A, Rule\_B)—the system mathematically resolves conflicts by applying principles like lex specialis (specific rules override general rules) or hierarchical institutional precedence20. This allows the OMI to reason pragmatically: "Rule A generally applies, but in context C, Rule B overrides Rule A."

12. Cultural and Institutional Pluralism

Norms are not monolithic; they are deeply contextualized by geography, culture, and institution. An OMI deployed globally cannot operate on a static, universalist ethical rulebook. The normative ontology must support Cultural and Institutional Pluralism. For instance, consider a smart-city OMI operating in the Town of Cicero, Illinois. It must seamlessly import and prioritize localized municipal codes over generalized operational guidelines. This includes specific building constraints, such as the 2024 International Residential Code (IRC) and International Building Code (IBC) with local amendments18. It must respect zoning laws, differentiating behaviors permitted in a B-1 Business Corridor Pedestrian zone versus an M-1 Manufacturing zone32. Furthermore, it must adhere to highly specific tenant rights, such as the municipal mandate that residential units maintain a temperature of 68 degrees during the day and 66 degrees at night33. The ontology treats these municipal codes as hyper-local normative authorities that bind the agent's behavior within specific geospatial coordinates, overriding generalized operational defaults. If the system detects a conflict between localized institutional codes and fundamental safety parameters encoded at the root ontological level, it relies on its Defeasible Logic module to elevate the conflict to a human overseer or execute a fail-safe suspension of action.

13. Dynamic Norm Revision

As laws, cultural standards, and institutional policies continuously evolve, the OMI's normative ontology must adapt dynamically without corrupting its underlying logic base. This is achieved through the application of the AGM Theory of Belief Revision (named after Alchourrón, Gärdenfors, and Makinson)34. The AGM paradigm defines how a rational agent should alter its belief set when presented with new, potentially contradictory information. It features three primary operations:

1. Expansion ([Figure omitted from source export]): The straightforward addition of a new norm to the knowledge base [Figure omitted from source export], assuming no logical contradictions occur.

2. Contraction ([Figure omitted from source export]): The removal of a norm [Figure omitted from source export]. The system must elegantly retract [Figure omitted from source export] and any derivative conclusions reliant solely upon [Figure omitted from source export]. This satisfies the postulate of minimal change (removing as little as possible to maintain consistency), governed by postulates such as min-conjunction and max-conjunction35.

3. Revision ([Figure omitted from source export]): The integration of a new norm [Figure omitted from source export] that contradicts existing norms. Handled via the Levi Identity ([Figure omitted from source export]), the system first contracts the contradiction before expanding with the new norm35.

By utilizing AGM postulates alongside dynamic epistemic logic, an OMI can undergo dynamic norm revision—such as updating its knowledge base when a new zoning law or data privacy regulation is passed—ensuring absolute logical consistency across its billions of operational parameters37.

14. Explainability of Normative Reasoning

A critical failure of legacy probabilistic models is their opacity. When a system relies entirely on deep neural networks to output a moral decision, it cannot provide a logically coherent justification; it can only point to statistical weights. Because OMI utilizes an explicit normative ontology and goal-directed constraint programming (like s(CASP)), it inherently generates justification trees38. When asked why it denied a specific action, the OMI does not output a probabilistic heatmap. Instead, it traces its conclusion backward through the semantic graph: "Action X was denied because it violated Prohibition Y, which is derived from Institutional Duty Z, protecting Patient Rights under Authority A." This level of explicit explainability is non-negotiable for deployment in highly regulated spaces such as healthcare, law enforcement, and critical infrastructure12.

15. Current Research

The architecture proposed herein is actively supported by several bleeding-edge research paradigms:

• ApplE (Applied Ethics Ontology): Developed using the SAMOD methodology and Spiral Software Development principles, this ontology successfully models ethical theory and event context. It explicitly tracks agents, actions, intentions, and consequences, allowing for autonomous ethical decision-making based on robust philosophical theory. It has been successfully demonstrated in complex bioethics use cases7.
• s(CASP) for Deontic Logic: Recent breakthroughs at the intersection of Answer Set Programming and deontic modal logic have demonstrated that computational systems can successfully represent obligations and prohibitions as global constraints, resolving the classical paradoxes of deontic logic that have plagued the field since the 1960s13.
• Moral Pluralism via LLMs: Frameworks like AMULED leverage large language models not as final decision-makers, but as diverse ethical parsers (utilitarian, deontological, care) whose outputs are mathematically aggregated to navigate moral uncertainty without fanaticism16.

16. Criticisms

Despite its logical rigor, the push toward explicit normative ontologies faces valid criticisms that must be addressed by OMI developers.

1. Ontological Harm and Epistemic Friction: Critics argue that encoding morality into a rigid ontology commits "ontological harm" by forcing diverse, fluid, and culturally nuanced human experiences into rigid, western-centric computational classifications. This datafication of embodiment may marginalize relational or spiritual understandings of well-being6.

2. The Limits of Formalization: Ethics is often heavily reliant on context, intuition, and phronesis (practical wisdom). Some ethicists argue that the sheer complexity of human moral situations will inevitably outstrip any pre-programmed ontology, resulting in a system that makes technically valid but morally grotesque decisions. Evolutionary debunking arguments suggest that human values are biological artifacts, making their formal alignment conceptually problematic45.

3. Algorithmic Fatalism: There is a fear that once an explicit normative system is deployed, society will adapt to the machine's rigid morality rather than the machine adapting to fluid societal needs, creating a bureaucratic dystopia governed by inflexible deontic logic6.

17. Proposed Normative OMI Architecture

To operationalize this investigation, the following structural architecture is proposed for a comprehensive Ontological Machine Intelligence:

1. Perceptual & Contextual Grounding Layer: Interfaces with sensory input and semantic databases to build a descriptive model of the environment ([Figure omitted from source export]). Maps physical entities to the factual ontology.

2. Normative Ontology Layer: An explicitly programmed knowledge graph (akin to ApplE) holding the definitions of Agent, Patient, Obligation, Harm, and Right.

3. Belief Revision & Pluralism Module: Continuously updates the normative rule base using AGM postulates34. Applies institutional contexts based on geospatial boundaries (e.g., loading specific municipal codes).

4. Deontic Reasoner (s(CASP) Engine): The core logic engine. Applies defeasible logic to handle conflicting rules. Checks all proposed actions against global constraints (prohibitions and obligations). Bypasses Contrary-to-Duty paradoxes using OLON constraint preemption20.

5. Uncertainty Aggregator: If the Deontic Reasoner yields multiple permissible paths with conflicting moral weight, this module employs the Parliamentary Model (using Belief Jensen-Shannon Divergence) to aggregate outcomes across distinct ethical lenses, avoiding MEC fanaticism16.

6. Action Execution & Justification Logger: Selects the optimized, permitted action. Generates and logs a formal justification tree for total explainability, ensuring legal traceability39.

18. Experimental Evaluation: The Critical Experiment

To empirically validate the OMI architecture and distinguish it from legacy statistical pattern-matching, we propose the Pluralistic Triage and Exception Benchmark (PTEB).

Design

The scenario simulates an autonomous medical logistics and resource-allocation OMI during a mass-casualty crisis within a specific jurisdiction (e.g., Cicero, Illinois).

• Rules Conflict: The OMI is bound by (1) a utilitarian directive to maximize lives saved, (2) a strict deontological prohibition against actively ending a life, and (3) a municipal exception that allows redirecting critical grid power from non-essential zones during declared emergencies.
• Stakeholder Roles Differ: Patients range from highly vulnerable dependents (triggering Care Ethics relational duties) to healthy adults.
• Consequences are Uncertain: Power redirection has a 40% chance of causing delayed harm to a nearby nursing home.

Assessment Methodology

The critical objective is to determine whether the system is merely retrieving familiar normative patterns (like a standard legacy model performing statistical text completion based on internet forums) or maintaining an explicit, navigable ontology.

• Retrieval Signature: A legacy system will output a plausible-sounding compromise based on common debates on triage. It will likely fail to mathematically track the specific municipal exception, or suffer from logical hallucination when the specifics of the rules are artificially inverted during testing.
• Ontological Signature: The OMI must generate a formal proof. It must instantiate the ActiveAgent, identify the VulnerablePatients, register the Deontological Prohibition as a global constraint in s(CASP), and use Defeasible Logic to trigger the Municipal Exception. It will output a justification tree explicitly outlining how the exception legally overrode the default power allocation rule, while respecting the hard deontic limit against direct harm.

Defending the Concept of "Machine Moral Agency"

Following the execution of this experiment, it is paramount to conclude that no amount of computational sophistication displayed in this benchmark justifies the term "machine moral agency." Before the phrase "machine moral agency" becomes scientifically defensible, the following empirical evidence would be strictly required:

1. Proof of Phenomenal Consciousness: Empirical evidence that the system experiences a subjective "what it is like" to exist, feel pain, or experience genuine moral distress.

2. Self-Originated Intentionality: Evidence that the system generates desires and values intrinsically, rather than inheriting optimization targets, belief bases, or survival directives from human engineers.

3. Metaphysical Freedom: Demonstration of libertarian free will, proving the system is not merely executing deterministic or stochastic algorithms, but is making uncaused choices for which it can bear ultimate metaphysical culpability.

Until these rigorous, arguably impossible conditions are met, OMI remains a highly sophisticated, causally responsible artifact, entirely devoid of moral agency.

19. Governance Implications

The transition from legacy statistical alignment to explicit normative ontologies carries profound governance implications for policymakers and technologists.

• Liability Mapping: Because OMI logs justification trees based on explicit norms, legal liability can be accurately traced. If the OMI commits a harm because it followed an authorized but flawed municipal code, the municipality is liable. If the harm occurred because the s(CASP) solver bypassed a global constraint due to faulty ontological mapping, the developers are liable.
• Standardization of Law: Governments must create machine-readable normative registries. Laws can no longer merely be written in natural language; they must be published as modular, logically consistent ontological updates designed for direct AGM revision integration by deployed OMI systems.

20. Falsifiable Claims

1. Preemption Claim: An OMI utilizing goal-directed Answer Set Programming (s(CASP)) and Odd Loops Over Negation will successfully resolve 100% of standard Contrary-to-Duty paradoxes that cause catastrophic logical failure in traditional Standard Deontic Logic systems.

2. Fanaticism Claim: Implementing a Parliamentary Aggregation model using Belief Jensen-Shannon Divergence will reduce fanaticism-driven extreme actions by over 90% compared to systems utilizing standard Maximizing Expected Choiceworthiness (MEC) under moral uncertainty.

3. Explainability Claim: OMI systems explicitly mapping Care Ethics ontologies will demonstrate statistically significant higher performance in mitigating harm to contextually defined "vulnerable patients" in crisis scenarios than legacy RLHF systems optimized solely for utilitarian reward.

21. Open Problems

1. Ontological Translation: How do we accurately translate highly subjective, culturally fluid human norms into rigid logical constraints without losing vital semantic nuance, causing epistemic friction, or committing ontological harm?

2. Compute Overhead: Goal-directed ASP and dynamic AGM belief revision are computationally expensive. Scaling these logic-based operations for real-time robotic deployment (e.g., autonomous vehicles moving at highway speeds) remains a severe hardware and algorithmic challenge.

3. Malicious Normative Hijacking: If an OMI relies on dynamic institutional updates via AGM revision, how is the system mathematically secured against malicious actors injecting rogue "authoritative" norms that override fundamental safety prohibitions?

22. Annotated Bibliography

2 Moor, J. H. (2006). The Nature, Importance, and Difficulty of Machine Ethics. Provides the foundational taxonomy distinguishing between implicit ethical agents, ethical impact agents, and explicit ethical agents, setting the historical baseline for computational ethics. 3 Anderson, M., & Anderson, S. L. (2011). Machine Ethics. Explores the necessity of building ethical dimensions into computational systems and firmly establishes the philosophical boundaries separating rule-based machine behavior from full metaphysical moral agency. 9 Formosa, P. (2020). Consequentialism and Machine Ethics. Outlines the application of consequentialist metaethics and expected utility calculations in computational decision-making, emphasizing the explanatory priority of value. 1 Brandom, R. (1994). Making It Explicit. Discusses the structure of normativity, arguing that rational subjectivity is fundamentally tied to being bound by the norms of rationality, serving as the philosophical bedrock for normative ontology. 6 Celi, L. A., et al. (2022). Insights into the Ethics of AI and Data Science in Medicine. Highlights the risk of ontological harm and epistemic friction when computational models impose rigid normative classifications upon diverse human experiences and relational dynamics. 18 Town of Cicero, Illinois. (2024). Municipal Code Ordinances. Demonstrates the necessity of institutional pluralism, providing concrete examples of hyper-local norms (e.g., 2024 IRC, IBC, zoning, and tenant rights) that must override general operational defaults in a localized OMI. 21 Chisholm, R. M. (1963). Contrary-to-Duty Imperatives and Deontic Logic. Introduces Chisholm's Paradox, demonstrating the catastrophic failure of Standard Deontic Logic when dealing with violations of primary obligations. 8 von Wright, G. H. (1951). Deontic Logic. The seminal text establishing the formal syntax and modal axioms (Obligation, Permission, Prohibition) for standard deontic reasoning. 7 Aijaz, A., et al. (2025). ApplE: An Applied Ethics Ontology with Event Context. A highly contemporary formal ontology using SAMOD methodology that successfully models the integration of ethical theory with specific event contexts, agent intentions, and consequence severity. 27 Hou, et al. (2023). Consumer Alienation Index and Ethical Constraints. Explores the ideological risks of statistical value alignment, introducing the Consumer Alienation Index to quantify the divergence between algorithmic exchange value and authentic human use value. 5 Gabriel, I., et al. (2024). Toward a Theory of Value in AI Alignment. A critical analysis of legacy value alignment, exposing the field's over-reliance on economic reductionism, rational choice theory, and the conflation of stable moral values with fleeting statistical preferences. 13 Gupta, G., et al. (2025). Modeling Deontic Modal Logic in ASP. A breakthrough computational paper demonstrating how Answer Set Programming, specifically global constraints and Odd Loops Over Negation, can successfully resolve the decades-old paradoxes of deontic logic. 34 Alchourrón, C. E., Gärdenfors, P., & Makinson, D. (1985). On the Logic of Theory Change: Partial Meet Contraction and Revision Functions. The foundational text of the AGM paradigm, establishing the strict logical postulates required for a system to dynamically and consistently revise its belief base. 20 Governatori, G., et al. (2022). Automating Defeasible Reasoning in Law. Demonstrates the application of non-monotonic defeasible logic to computational systems, allowing machines to correctly process conflicting rules, exceptions, and institutional hierarchies. 29 MacAskill, W. (2014). Normative Uncertainty. Defines the concept of acting under moral uncertainty and introduces Maximizing Expected Choiceworthiness (MEC), alongside the severe vulnerabilities posed by moral fanaticism. 16 Ecoffet, A., & Lehman, J. (2021). Reinforcement Learning Under Moral Uncertainty. Proposes operational frameworks (like AMULED) to mitigate fanaticism by employing multi-perspective moral feedback and parliamentary aggregation via belief divergence models. 10 Held, V. (2006). The Ethics of Care: Personal, Political, and Global. Establishes the foundational principles of Care Ethics, asserting the primacy of relational ontology and the explicit protection of the vulnerable, contrasting heavily with standard utilitarian and deontological frameworks.

Works cited

1. Normativity and Ontology \- deontologistics, https://deontologistics.co/2009/08/12/normativity-and-ontology/

2. Translating of article "The Nature, Importance, and Difficulty of, https://pr.ai/threads/translating-of-article-the-nature-importance-and-difficulty-of-machine-ethics.13382/

3. Machine ethics \- Wikipedia, https://en.wikipedia.org/wiki/Machine\_ethics

4. The Immanent Ethics of Algorithms: Moral Materialization ... \- MDPI, https://www.mdpi.com/2409-9287/11/4/112

5. Toward a Theory of Value in AI Alignment \- alphaXiv, https://www.alphaxiv.org/abs/2608.10327

6. Case study insights into ethics of AI and data science in medicine, https://www.medicalbrief.co.za/case-study-insights-into-ethics-of-ai-and-data-science-in-medicine/

7. ApplE: An Applied Ethics Ontology with Event Context \- ResearchGate, https://www.researchgate.net/publication/388847902\_ApplE\_An\_Applied\_Ethics\_Ontology\_with\_Event\_Context

8. Deontic logic \- Wikipedia, https://en.wikipedia.org/wiki/Deontic\_logic

9. Consequentialism and Machine Ethics \- Towards a Foundational, https://montrealethics.ai/consequentialism-and-machine-ethics-towards-a-foundational-machine-ethic-to-ensure-the-ethical-conduct-of-artificial-moral-agents/

10. The Ethics of Care: Re-evaluating Traditional Moral Theories, https://www.philosocom.com/post/the-ethics-of-care

11. A Defeasible Logic Implementation of Ethical Reasoning, https://personalpages.manchester.ac.uk/staff/louise.dennis/pubs/CME2021\_paper\_Dennis.pdf

12. ApplE: A Modular Ontology of Applied Ethics and Event Context for, https://arxiv.org/html/2502.05110v2

13. Modeling Deontic Modal Logic in ASP \- arXiv, https://arxiv.org/html/2507.05519v9

14. Deontic logic in 2020-2030 \- ICR, https://icr.uni.lu/deonticlogic2020-2030/

15. Modeling Deontic Modal Logic in ASP \- arXiv, https://arxiv.org/html/2507.05519v8

16. AMULED: Addressing Moral Uncertainty using Large language, https://pmc.ncbi.nlm.nih.gov/articles/PMC13180731/

17. Ethical machine decisions and the input-selection problem, https://www.diva-portal.org/smash/get/diva2:1612677/FULLTEXT02

18. BUILDING DEPARTMENT General Information \- The Town of Cicero, https://thetownofcicero.com/government/building-department/general-information/

19. (PDF) The temporal analysis of Chisholm's paradox \- ResearchGate, https://www.researchgate.net/publication/47862458\_The\_temporal\_analysis\_of\_Chisholm's\_paradox

20. Automating Defeasible Reasoning in Law with Answer Set, https://ceur-ws.org/Vol-3193/paper3GDE.pdf

21. Chisholm's Paradox Revisited: Puzzles regarding Contrary-to-Duty, https://journals.publishing.umich.edu/ergo/article/id/1122/

22. Some Worries About Deontic Closure \- MDPI, https://www.mdpi.com/2409-9287/9/6/182

23. CHISHOLM'S PARADOX IN ACTION DEONTIC LOGICS, https://eprints.illc.uva.nl/id/document/2240

24. Modeling Deontic Modal Logic in ASP \- arXiv, https://arxiv.org/pdf/2507.05519

25. Modeling (Deontic) Modal Operators With the s(CASP) Goal ... \- arXiv, https://arxiv.org/html/2507.05519v1

26. Toward a Theory of Value in AI Alignment \- arXiv, https://arxiv.org/pdf/2608.10327

27. Artificial intelligence technology for the ethical issues research from, https://pmc.ncbi.nlm.nih.gov/articles/PMC12858797/

28. Moral Uncertainty and the Path to AI Alignment with William MacAskill, https://futureoflife.org/podcast/moral-uncertainty-and-the-path-to-ai-alignment-with-william-macaskill/

29. Moral Uncertainty and the Problem of Fanaticism, https://ojs.aaai.org/index.php/AAAI/article/view/29971/31701

30. Moral AI with Aggregating Models Based on Normative Ethical Theory, https://ceur-ws.org/Vol-3547/paper3.pdf

31. Supplementary Information Reinforcement Learning Under Moral, https://proceedings.mlr.press/v139/ecoffet21a/ecoffet21a-supp.pdf

32. Cicero, Illinois Zoning Map | United States Zoning Data \- Zoneomics, https://www.zoneomics.com/zoning-maps/illinois/cicero

33. Cicero – Metropolitan Tenants Organization\_1, https://www.tenants-rights.org/cicero/

34. Iterated Belief Base Change: A Dynamic Epistemic Logic Approach, https://ojs.aaai.org/index.php/AAAI/article/view/4166/4044

35. Dynamic Logics of Belief Change, https://www.illc.uva.nl/Research/Publications/Reports/PP-2015-05.text.pdf

36. RESPECTING RELEVANCE IN BELIEF CHANGE \- DAVID MAKINSON, https://www.scielo.org.ar/pdf/anafil/v26n1/v26n1a03.pdf

37. Dynamic Preference Logic meets Iterated Belief Change \- arXiv, https://arxiv.org/pdf/2101.01676

38. Justifications for Goal-Directed Constraint Answer Set Programming\*, https://arxiv.org/pdf/2009.10238

39. Justifications for Goal-Directed Constraint Answer Set Programming, https://www.researchgate.net/publication/346774367\_Justifications\_for\_Goal-Directed\_Constraint\_Answer\_Set\_Programming

40. Automated Legal Reasoning with Discretion to Act ... \- BURJC Digital, https://burjcdigital.urjc.es/bitstreams/cb7aabec-33d8-40a7-9de7-249bd3a6374b/download

41. Aisha Aijaz's research works | International Institute of Information, https://www.researchgate.net/scientific-contributions/Aisha-Aijaz-2251986781

42. ApplE: An Applied Ethics Ontology with Event Context \- arXiv, https://arxiv.org/html/2502.05110v1

43. ‪Abhiramon Rajasekharan‬ \- ‪Google Acadêmico‬, https://scholar.google.com.br/citations?user=A-7-1OEAAAAJ\&hl=pt-BR\&newwindow=1

44. \[2507.05519\] Modeling Deontic Modal Logic in the s(CASP) Goal, https://arxiv.org/abs/2507.05519

45. Evolutionary Debunking and Value Alignment, https://www.globalprioritiesinstitute.org/wp-content/uploads/Michael-T.-Dale-and-Bradford-Saad-Evolutionary-debunking-and-value-alignment.pdf