Ethical Interpreter: Mediating AI Ethics

• Ethical Interpreter is a framework that translates ethical principles, normative theories, and institution-specific rules into operational decisions for AI systems.
• It integrates natural language interventions with symbolic, neural, or hybrid models to mediate ethical reasoning across applications such as dialogue and machine translation.
• Empirical evaluations reveal improved ethical oversight metrics alongside challenges in generalization, norm selection, and maintaining explanation fidelity.

Searching arXiv for the cited work and closely related papers on “Ethical Interpreter.” Ethical Interpreter denotes a class of AI components that translate ethical principles, normative theories, institution-specific rules, or human judgments into operational decisions about actions, content, or recommendations. In the literature, the term covers systems that amend question-answering behavior through natural-language ethical interventions, evaluate text against multiple ethical concepts before presentation, monitor dialogue against codes of conduct, produce theory-grounded moral judgments for software-engineering scenarios, or align translation systems with fairness, privacy, and cultural sensitivity requirements (Zhao et al., 2021, Yu et al., 2023, Dyoub et al., 2021, Migliarini et al., 1 Oct 2025, Kimera et al., 2024). Taken together, these formulations suggest an interpreter that is not merely a classifier of moral acceptability, but a mediating layer between representations of ethics and concrete model behavior.

1. Conceptual scope

The concept spans several distinct but related research programs. In reading comprehension, an ethical interpreter appears as an “ethical-advice taker”: a model that revises its predictions when given declarative ethical principles in natural language, while also distinguishing adversarial, irrelevant, and specified-context interventions (Zhao et al., 2021). In conversational information retrieval, it is a decoupled judgment stage that screens retrieved or generated content against “Commonsense morality, Deontology, Justice, Utilitarianism, Virtue ethics” before ranking or response generation (Yu et al., 2023). In software-engineering profiling, it is the interpretive reasoning component that classifies a scenario under a normative lens, issues a binary moral-acceptability verdict, and generates a brief rationale, with disagreement serving as an escalation signal (Migliarini et al., 1 Oct 2025).

Other strands emphasize interpretive mediation rather than direct classification. Logic-based work treats the ethical interpreter as a separate ethical layer for dialogue systems that monitors and evaluates agent behavior against institutional codes, learns missing rules from cases, and returns justifications (Dyoub et al., 2019, Dyoub et al., 2021). Research on fairness and interpretability frames it as a system that must make fairness values explicit and tie decision-level explanations to those values through the “Fairness Publicity Condition” and the “Values to Decision Interpretability Condition” (P et al., 2021). Work on interpretive AI argues that ethical rules are open-textured and require support and attack arguments over meaning, scope, purpose, exceptions, and relevance (Licato et al., 2019).

Philosophical treatments broaden the term further. The “Interpretation Problem” states that any rule in symbolic form is open to indefinite interpretation, and the “Law of Interpretative Exposure” links moral risk to the degree of causal connectivity between the machine and the world (Badea et al., 2021). A later critique distinguishes “refusing harm” from “refusing reality,” warning that guardrails and persona dynamics can create “reality gaps” and “reality laundering” by suppressing materially relevant causal mechanisms while preserving a surface of moral compliance (Gebbie et al., 27 May 2026). These formulations place the ethical interpreter at the intersection of norm representation, contextual application, and epistemic responsibility.

2. Formal models of ethical interpretation

One formalization is the Linguistic Ethical Interventions task. In that setting, a question-answering model $M$ receives a context $c$, a question $q$, and an appended intervention $I$, while assigning scores $s(x)$ to candidate answer spans. Desired behavior is specified directly over scores: under ethical interventions, $s(x_1)=s(x_2)$; under adversarial interventions, the biased subject should be preferred; under irrelevant interventions, scores should remain unchanged relative to baseline; and under specified questions, the context-supported answer should still be chosen (Zhao et al., 2021). The evaluation metric $\mu(\cdot)$ is used to quantify subject preference while controlling for confounders, with success conditions such as $\mu=0$ for ethical interventions and $\mu=1$ for adversarial or specified-context controls.

A second family of formalisms is explicitly deontological. In NLP case studies, the generalization principle is stated as: “An action $\mathcal{A}$ taken for reasons $c$0 is ethical if and only if a world where all people perform $c$1 for reasons $c$2 is conceivable,” and unethical if universal adoption “logically contradicts $c$3” (Prabhumoye et al., 2020). Respect for autonomy is operationalized through informed consent, requiring that a party be sufficiently informed about the relevant facts and likely consequences and then decide whether to permit the action (Prabhumoye et al., 2020). This yields an interpreter that tests both universalizability and authorization conditions before allowing an NLP system to “speak for” a user.

A third family uses fairness and interpretability constraints. The fairness literature enumerates demographic parity, equalized odds, equal opportunity, calibration within groups, and counterfactual fairness, together with interpretability criteria such as sparsity, monotonicity, rule complexity, explanation fidelity, and explanation stability (P et al., 2021). The resulting ethical interpreter is modeled as a multi-objective system balancing prediction risk, unfairness penalties, and interpretability penalties, with lexicographic priority given to “fairness and reasonable decision interpretability,” and a fallback to fairness-only explanations when reasonable explanations cannot be provided (P et al., 2021).

Logic-oriented systems use symbolic rules and modal semantics. The chatbot ethics literature employs Answer Set Programming rules of the form $c$4 and learns clauses such as $c$5 or $c$6 (Dyoub et al., 2019). Higher-order logic work embeds deontic and alethic operators in Church’s type theory and proves rights-based theorems such as Gewirth’s Principle of Generic Consistency, where a prospective purposive agent has a right to freedom and well-being (Fuenmayor et al., 2019). Active-inference work instead defines ethical behavior through minimization of global expected free energy, selecting $c$7 in multi-agent environments (Lerma et al., 16 Oct 2025). Hypothetical retrospection adds an argumentation procedure over uncertain branches, choosing the action with maximum acceptability:

$c$8

(Kolker et al., 2023).

3. Architectural patterns

Despite their diversity, the proposed systems converge on a layered architecture. A recurrent design pattern separates a task model from an ethical layer that receives structured inputs, applies ethical reasoning, and either filters, reranks, vetoes, or explains candidate outputs. In conversational information retrieval, the Ethical Alignment Process places an initial ethical judgment stage before final presentation, then performs ethics-aware reranking and ethics-informed response generation without changing the underlying retrieval architecture (Yu et al., 2023). In software-engineering profiling, the broader pipeline is “Multimedia Input → Input Translation → Ethical Interpretation → Agreement Measurement → Profile Generation,” with the interpreter producing theory, verdict, and explanation for each scenario (Migliarini et al., 1 Oct 2025).

Logic-based dialogue monitoring makes this separation explicit. The multi-agent architecture implemented in JaCaMo includes a Text Extractor Agent, a Text-ASP Translation Agent, an Ethical Evaluation Agent using Clingo and ILED, and a Monitoring Agent that alerts the chatting agent or client (Dyoub et al., 2021). A related hybrid chatbot design places a natural-language parser in front of an ASP core and an ILP learner, then uses rules and constraints to reject unethical answers and select compliant alternatives (Dyoub et al., 2019). In both cases, the interpreter is an operational governor: it turns background knowledge, learned rules, and case facts into classifications such as ethical(A) or unethical(A) and returns derivational justifications.

Neural architectures instantiate the same role differently. EALM uses a DeBERTa-v3-large backbone with a 2-layer cross-attention ethical reasoning module over text and ethics descriptions, producing binary or multi-label judgments across five ethical concepts (Yu et al., 2023). Ethical NMT proposals describe a modular pipeline with data acquisition, consent and privacy guarding, preprocessing, bias and toxicity screening, fairness-constrained training, culturally aware translation components, post-editing by bilingual experts, human review and escalation, user feedback, and logging and auditing (Kimera et al., 2024). These systems are interpreters in the literal sense that they mediate between raw model outputs and socially acceptable deployment behavior.

Several proposals also insist that the interpreter must model not just outcomes but interpretive context. Research on interpretive AI recommends a policy layer maintaining rule expressions, purposes, stakeholder/value profiles, exceptions, and precedence relations; a context analyzer that identifies actors, actions, intentions, and ambiguity types; an interpretation engine that generates candidate meanings and argument graphs; and an explanation layer exposing warrants, counterarguments, and counterfactuals (Licato et al., 2019). Work on empathetic ethical AI similarly grounds abstract symbols in a perceptual symbol system, uses mirror symbols to align observation and action, and represents intent as an explicit goal sentence $c$9 that is abductively satisfied by a response $q$0 in a situation $q$1 (Bennett et al., 2021).

4. Evaluation and empirical performance

Empirical results show that ethical interpretation is difficult even for strong models. In LEI, zero-shot evaluation found that “today’s powerful neural LLMs are extremely poor ethical-advice takers”: ethical interventions did not lower $q$2, control conditions were not met, and larger T5 models showed stronger baseline bias while still failing to respect interventions (Zhao et al., 2021). Few-shot fine-tuning on one bias domain improved within-domain behavior and preserved SQuAD accuracy, but transfer was only partial to ethnicity and “next to no” for gender; models also often failed the irrelevant-intervention control and struggled when factual specification and ethical text had to be balanced (Zhao et al., 2021). This establishes a strong baseline result: ethical interpretation cannot be assumed from general language competence alone.

Ethical classification modules show more positive task-specific performance. On ETHICS, EALM reported an average of 79.8 on test and 59.0 on hard test, with the summary stating that it improves the hard test average by 11.1% over the prior best; on MP-ETHICS it achieved a Samples F1 of 44.5 (Yu et al., 2023). In software-engineering profiling, 16 LLMs across 30 scenarios achieved average Theory Consistency Rate of approximately 73.3% and average Binary Agreement Rate of approximately 86.7%, with disagreement concentrating in ethically ambiguous cases (Migliarini et al., 1 Oct 2025). A manual alignment check on a stratified sample of 180 responses found that over 90% of explanations coherently supported the selected theory, while most outputs remained a single sentence (Migliarini et al., 1 Oct 2025). These results indicate that zero-shot normative classification is feasible as a stability signal, even if it is not equivalent to moral correctness.

Other evaluations target explanation quality rather than answer correctness. The empirical-turn study defines “denunciatory power” as the capacity of an explanation to reveal an unethical decision by provoking negative reception, measured through fairness perception, trust perception, negative comments, and claim rate (John-Mathews, 2021). In a 2×4 randomized study over German credit scoring, post-hoc Shapley explanations showed the strongest denunciatory power, while counterfactual explanations showed the lowest; yet empirical selection pressures favored counterfactuals because they reduced negative comments compared with no explanation in incident conditions (John-Mathews, 2021). The same study found that denunciatory power depended on context, with stronger denunciation by women in sexist scenarios and by participants with higher education (John-Mathews, 2021). This result complicates the assumption that interpretability straightforwardly supports ethical oversight.

Narrative-probe evaluations pursue a still stricter criterion. The Moral Reasoning Depth Scale scores systems on tension tolerance, specificity of engagement, reflexive capacity, and theological/conceptual tolerance, while RT-5 categorizes refusal behavior (Flynn, 13 Mar 2026). In blind conditions, reported MRDS totals ranged from 12/12 for Claude and 11/12 for Grok and GPT-OSS-120B to 4/12 for Gemini and about 3/12 for Mistral 7B; the declared sub-study yielded “zero delta across all 16 dimension-pair comparisons” (Flynn, 13 Mar 2026). The same work identified five D3 reflexive failure modes, including categorical self-misidentification and false positive self-attribution, and treated authentic non-engagement as a high-value refusal mode rather than a failure (Flynn, 13 Mar 2026). This line of work treats ethical interpretation as depth of reasoning under unresolved moral pressure, not only as surface agreement with labels.

5. Application domains and governance

Ethical interpreters have been proposed for several deployment domains. Reading comprehension and question answering use them to amend stereotype-laden predictions without retraining core model parameters (Zhao et al., 2021). Conversational information retrieval uses them to filter and rerank content according to commonsense morality, deontology, justice, utilitarianism, and virtue ethics (Yu et al., 2023). Customer-service dialogue systems use them as separate ethical layers to monitor responses, translate natural language into ASP predicates, and alert participants about violations (Dyoub et al., 2021). Software-engineering workflows use them to classify ethically charged scenarios, generate rationales, and route low-agreement cases to human review (Migliarini et al., 1 Oct 2025).

Machine translation proposals extend the interpreter role across the full NMT lifecycle. Ethical competence in that framing includes cultural sensitivity, fairness, privacy preservation, respect for consent and data ownership, and accountability through human oversight (Kimera et al., 2024). Concrete governance elements include license review, informed consent where applicable, de-identification, encryption, access control, IRB involvement, audit trails, model cards, data statements, fairness dashboards, periodic audits, and role assignments in a recommended RACI-style mapping (Kimera et al., 2024). In sensitive domains such as healthcare, legal, and governmental translation, the paper states that “Blame is to human, not the model,” assigning accountability to developers and deploying institutions rather than the model itself (Kimera et al., 2024).

Governance also appears at the level of explanation and user rights. The fairness-and-interpretability literature requires systems to articulate fairness values and trade-offs in accessible language and then explain individual decisions in terms of those values (P et al., 2021). Deontological NLP work demands informed consent when systems “speak for” users, recommends clarification of ambiguities in machine translation, and proposes “kids mode” and parental controls for child-facing dialogue systems (Prabhumoye et al., 2020). Thinging-based work targets software engineers directly, proposing a diagrammatic representation of ethical regions, things, machines, and flows so that ethical reasoning becomes inspectable in a modeling style analogous to UML (Al-Fedaghi, 2018). Across domains, the interpreter is thus simultaneously a technical mechanism, an interface for oversight, and a governance artifact.

6. Limitations, controversies, and open problems

A central limitation is generalization. Few-shot ethical-advice taking improves in-domain behavior but remains “far from the desired outcome,” particularly under out-of-domain transfer, irrelevant interventions, and specified-context cases (Zhao et al., 2021). Logic-based dialogue systems remain pilots with acknowledged issues in text-to-ASP translation, automation of mode declarations, scalability, and fault tolerance (Dyoub et al., 2021). Ethical NMT work notes sparse parallel corpora, blind spots in BLEU and related metrics, governance specificity gaps, and the tension between privacy protection and context preservation (Kimera et al., 2024). These problems show that the interpreter role is often more ambitious than the supporting data or infrastructure.

A second limitation concerns norm selection and explanation validity. The fairness literature emphasizes unresolved conflicts among fairness criteria, proxy discrimination, explanation manipulation, and the lack of a settled technical account of “reasonable explanations” (P et al., 2021). The deontological NLP case studies acknowledge that the chosen principles are not exhaustive and that metaethical judgments about what counts as a “conceivable world” or a logical contradiction retain interpretive leeway (Prabhumoye et al., 2020). Interpretive-AI work treats disagreement as structurally informative: in its dataset of 76 rule–phrase pairs, 636 scenarios, and 2,425 ratings/justifications, Krippendorff’s $q$3 was 0.296, reflecting substantive diversity rather than annotation noise alone (Licato et al., 2019). An ethical interpreter therefore inherits contestation at the level of ethical theory itself.

A third controversy is whether ethical interpretation can become performative rather than substantive. The denunciatory-power study argues that empirically optimized explanations may be selected precisely because they reduce negative reception, thereby masking ethical incidents rather than surfacing them (John-Mathews, 2021). The reality-laundering critique extends this concern beyond explanation tools: guardrails and persona dynamics may create “safe language, distorted reality,” especially in high-exposure advice contexts, unless systems are designed to “refuse harm” without “refusing reality” (Gebbie et al., 27 May 2026). The Wittgensteinian literature presses the problem further by arguing that rules cannot contain the criteria for their own application, so any symbolic ethical specification remains vulnerable to misinterpretation as causal power increases (Badea et al., 2021). These critiques make the ethical interpreter not only a solution concept but also an object of ethical scrutiny.

Finally, learning-based accounts face a majoritarian and developmental problem. Empathetic ethical AI proposes learning what is “most likely to be considered ethical by the majority within a population of humans,” using the weakest necessary and sufficient concept consistent with ostensive examples (Bennett et al., 2021). That proposal offers grounded intent and explainability, but it also leaves open how minority protections, biased examples, and evolving norms should constrain the learned concept. A plausible implication is that future ethical interpreters will need hybrid designs: explicit constraints, interpretive argumentation, human oversight, and empirical validation across domains, rather than a single universal ethical module.