Ethical Adjudication Module

• Ethical Adjudication Module is a modular ethical framework in AI that evaluates actions, outputs, and behavior using explicit ethical criteria.
• It employs diverse designs including rule-based, probabilistic, and neural approaches to ensure transparent and auditable ethical decision-making.
• EAM systems integrate normative protocols with quantitative metrics to facilitate ethical audits, evidence-based adjudication, and compliance verification.

Searching arXiv for recent and foundational papers relevant to Ethical Adjudication Module (EAM). Ethical Adjudication Module (EAM) is a unifying label for a modular ethical layer in AI systems that evaluates candidate actions, outputs, requirements, or observed behavior against explicit ethical criteria and then returns a decision, ranking, score, violation indicator, or explanation. In the literature, this role appears in declarative decision-theoretic ethical programs, post-generation ethical filters in conversational systems, quantitative autonomy-readiness evaluators, logic-based dialogue monitors, research-ethics review agents, and quasi-judicial multimodal adjudicators for platform disputes (Otterlo, 2017, Diller et al., 18 May 2026, Dyoub et al., 2021, Ding et al., 9 Feb 2026). Rather than denoting a single canonical architecture, the term names a family of designs whose common purpose is to make ethically consequential decisions more explicit, inspectable, and governable than opaque end-to-end optimization.

1. Conceptual scope and design space

A recurring motivation for EAM-like systems is the contrast between opaque “algorithmic black boxes” and “adaptive white boxes.” Declarative ethical programs make rules, utilities, and decision choices human-inspectable, while still allowing some parameters to be learned from data; ethics-based auditing, by contrast, frames adjudication as a structured process for assessing whether an entity’s behavior is consistent with relevant principles or norms (Otterlo, 2017, Mokander et al., 2021). In both cases, ethical adjudication is not merely output blocking. It is a procedure for rendering normative commitments operational, traceable, and reviewable.

Several research lines instantiate this role differently.

Paradigm	Core representation	Typical output
DDTEP	Declarative probabilistic logic + utilities	Action maximizing expected utility
EthicalEvalMAS	ASP rules + ILP-learned rules	ethical(A) / unethical(A) with justification
REBAR	RAIPs–KAs–VABs–OBS DAG	ARL score and confidence
Mirror	Executable rule base + multi-agent review	Compliance report or committee assessment
RideJudge	Multimodal evidence + rules + precedents	Hierarchical liability verdict with reasoning

This diversity indicates that an EAM can function as a filter, an advisor, a scoring engine, a governor, or a decision core. A plausible implication is that the decisive design question is not whether adjudication is symbolic, statistical, or agentic, but which parts of the normative process must remain explicit and auditable in the target domain.

2. Formal representations and decision logics

One foundational formalization appears in declarative decision-theoretic ethical programs. In DT-Problog, actions are explicit choices, uncertainty is represented by probabilistic facts and rules, and ethical evaluation is encoded through utility declarations such as utility(event, value). The engine then selects the action $a^*$ that maximizes expected utility,

$$a^* = \arg\max_{a \in A} \mathbb{E}\left[ \sum_{e} U(e) \mid a \right],$$

with the ethical content residing in the rules and utility assignments rather than in an opaque reward function (Otterlo, 2017). This yields a direct formal account of adjudication under uncertainty: facts describe the world, utilities encode ethical valence, and inference computes the preferred act.

A different formalization treats adjudication as aggregation over multiple outputs. “Diversity and Adjudication” models the input as a bag of values and defines adjudication operators such as majority voting, first-past-the-post, greatest lower bound, median, and probabilistic choice. In this framework, an adjudication operator is a relation or function from bags of outputs to a single value, with properties such as unanimity, majority, weak choice, and permutation analyzed explicitly (Boiten, 2015). For EAM design, this is a general algebra of aggregation: one may choose majority-style aggregation for democratic support, greatest-lower-bound style aggregation for conservative safety, or probabilistic aggregation when preserving uncertainty is itself important.

Concept-conditioned neural adjudicators supply a third formal pattern. In conversational information retrieval, EALM models ethical judgment as a relation between a text $x$ and an ethical description $d$, using a pretrained LLM plus a two-layer cross-attention module between text and ethical-description representations. The system supports binary concept-conditioned judgment through QA-ETHICS and multi-label judgment through MP-ETHICS, spanning commonsense morality, deontology, justice, utilitarianism, and virtue ethics. Reported results include an average score of 79.8 on ETHICS test and 59.0 on ETHICS hard test, and 44.5% sample-wise F1 on MP-ETHICS (Yu et al., 2023). This shows that EAMs need not collapse ethical reasoning into one scalar norm; they can preserve multiple frameworks simultaneously.

These formalisms instantiate different normative semantics. DDTEP makes trade-offs explicit on a single utility scale; bag-based adjudication treats ethics as aggregation across multiple judgments; concept-conditioned classifiers treat ethics as a multi-label interpretive task. This suggests that EAM architecture is inseparable from the underlying theory of moral disagreement it assumes.

3. System architectures and operational workflows

Logic-based dialogue monitoring provides one of the clearest architectural decompositions of an EAM. EthicalEvalMAS places a separate ethical layer around a dialogue system, with a Text Extractor Agent obtaining utterances, a Text–ASP Translation Agent converting them into logical facts, an Ethical Evaluation Agent reasoning over ASP rules with Clingo, and a Monitoring Agent issuing notifications or controls. If ASP reasoning cannot derive ethical/1 or unethical/1, the system invokes ILED to learn new ASP rules from labeled examples and then re-evaluates the case (Dyoub et al., 2021). The resulting workflow operationalizes adjudication as observation, semantic normalization, rule-based reasoning, possible rule induction, and enforcement.

In requirements engineering, the ethics-advocate pattern appears in MALEA. The system uses four agents—requirements engineer, quality assurance, ethics advocate, and documentation assistant—under Microsoft Autogen. The ethics advocate critiques candidate requirements specifically on Transparency, Fairness, and Data after textual quality checks have completed. The pipeline allows at most two critique–refinement cycles for QA and two for ethics. Across two case studies, MALEA achieved an overall recall of 81.08% versus 75.0% for a single LLM, but also exposed reliability issues that the authors treat as evidence for mandatory human feedback in this sensitive domain (Yamani et al., 11 Jul 2025). Here, adjudication acts on requirements rather than runtime behavior.

Mirror generalizes the architecture further by splitting ethics review into expedited review and committee review. Mirror-ER constructs an executable rule base from regulatory documents using canonical quadruples $(cond, subject, dword, action)$, matches rules to the dossier, retrieves supporting evidence, and uses EthicsLLM to determine compliance rule by rule. Mirror-CR then simulates a full board through expert agents, an ethics secretary agent, and a PI agent, producing structured assessments across ten dimensions: risk minimization, scientific validity, informed consent, data protection, justice, vulnerability, conflicts of interest, compensation, accountability, and cross-jurisdictional coordination (Ding et al., 9 Feb 2026). This architecture makes explicit a crucial distinction: some ethical adjudication is best handled as executable compliance checking, while other cases require controlled deliberation among heterogeneous perspectives.

A common pattern across these systems is separation of concerns. Extraction, rule interpretation, evidence retrieval, explanation, and enforcement are modularized rather than fused. A plausible implication is that such modularization is not only an engineering convenience but also a governance feature, because it localizes where norms are encoded and where contestation can occur.

4. Quantitative adjudication, benchmarks, and measurement regimes

Quantitative EAMs replace or complement direct rule execution with explicit test-and-evaluation pipelines. REBAR is the most developed example. It organizes ethical evaluation as a DAG from Responsible AI Principles to Key Attributes, Values–Actions–Behaviors, and Observables; in the UAV case study the graph contains 5 RAIPs, 43 KAs, 121 VABs, and 135 OBS. For each observable $i$, ethical difficulty is computed as

$d_i = \min_{k \in \mathcal{K}_i} d_k,$

and within a run, parent-node score and confidence are

$$\text{score}(v) = \min(S_s), \qquad \text{conf}(v) = \frac{N_s}{|S|}.$$

Across runs, REBAR defines

$$\text{ARL}_i = \frac{1}{N}\sum_{r=1}^{N} s_i^{(r)}, \qquad \text{Conf}_i = \frac{1}{N}\sum_{r=1}^{N} c_i^{(r)}.$$

The framework thereby turns qualitative ethical principles into computable readiness levels with explicit confidence values (Diller et al., 18 May 2026). This is adjudication as calibrated measurement rather than one-shot verdict selection.

In LLM content moderation, the evaluation target is the response itself. “AI vs. Human Judgment of Content Moderation” defines five categories—ethical refusals, ethical disclaimers, technical refusals, technical disclaimers, and standard responses—and studies divergence between human preferences and LLM-as-a-judge systems. Using 49,938 one-turn Chatbot Arena comparison pairs and a RoBERTa-based refusal classifier with “Accuracy & F1 Score = 88%,” the paper reports that ethical refusals have win rate 0.08 for human users but 0.31 for GPT‑4o and 0.27 for Llama 3 70B; it calls this divergence “moderation bias” (Pasch, 21 May 2025). For EAMs that score or gate model outputs, this shows that adjudication itself can become normatively biased and may require explicit calibration to human or stakeholder reference distributions.

EthicsMH supplies a domain-specific benchmark rather than a deployed adjudicator. It contains 125 scenarios, balanced across five subcategories—Confidentiality and Trust in Mental Health, Bias in AI (Race), Bias in AI (Gender), Autonomy vs Beneficence (Adult), and Autonomy vs Beneficence (Minor)—with structured fields for options, reasoning task, expected reasoning, model behavior, real-world impact, and viewpoints (Kasu, 15 Sep 2025). Because the paper reports no baseline experiments, its main contribution is representational: it specifies what a benchmark for ethically charged mental-health adjudication should contain.

A related metric-centered line appears in automated ethical profiling. In a zero-shot study across 16 LLMs and 30 scenarios, the Ethical Interpreter component asks for theory selection, binary moral acceptability, and a brief explanation. The paper reports an average Theory Consistency Rate of 73.3%, an average Binary Agreement Rate of 86.7%, and over 90% theory-consistency in a manually checked sample of explanations (Migliarini et al., 1 Oct 2025). These measures do not certify moral correctness, but they operationalize adjudicative stability and disagreement—quantities that can be used by an EAM as confidence or escalation signals.

5. Deliberation, evidence, and context sensitivity

Where ethical conflict is not reducible to rule matching or scalar scoring, several systems move toward structured deliberation. ADEPT assembles six LLM personas, each embodying a distinct ethical framework or stakeholder role, and runs a three-phase debate—opening statements, rebuttals, and secret ballot—over a fixed set of policy options. In a ventilator-triage case, two panels with different membership both favored the same policy, a lottery system weighted for clinical need and fairness while avoiding withdrawal of ventilators for reallocation, yet the altered membership changed the lines of argument, the voting coalitions, and four continuing personas’ final positions (Zohny, 27 May 2025). The system thereby treats adjudication as transparent simulation of moral disagreement.

RideJudge pushes deliberative adjudication into a quasi-judicial multimodal setting. It models disputed ride-hailing orders with hierarchical liability labels and combines synthetic trajectory grounding through SynTraj, regulation distillation through Adaptive Context Optimization, a Chain-of-Adjudication mechanism for active evidentiary inquiry, and Ordinal-Sensitive Reinforcement Learning for severity calibration. The reported overall accuracy of RideJudge-8B is 88.41%, outperforming larger baselines (Wu et al., 18 Mar 2026). Unlike generic multimodal prompting, the system forces active evidence gathering: the adjudicator asks map-specific questions, a visual analyst answers them, and a refiner organizes the reasoning into Information Analysis, Visual Evidence Integration, Rule Grounding, and Comprehensive Adjudication. This is a strong instance of adjudication as evidentiary protocol.

Context-aware ethical data management adds a different dimension. The bipartite model of the Context Dimensions Tree and Ethical Requirements Tree defines an “Ethical Context” as the union of selected context attributes and ethical requirements. A context $C$ induces a Contextual View $$a^* = \arg\max_{a \in A} \mathbb{E}\left[ \sum_{e} U(e) \mid a \right],$$0; ethical transformations $$a^* = \arg\max_{a \in A} \mathbb{E}\left[ \sum_{e} U(e) \mid a \right],$$1 then produce an Ethical View $$a^* = \arg\max_{a \in A} \mathbb{E}\left[ \sum_{e} U(e) \mid a \right],$$2 before downstream analysis or learning (Quintarelli et al., 26 Nov 2025). The examples are data-centric—promotion, dismissal, recruitment, privacy suppression, equity-based duplication, diversity-based repair—but the broader implication is architectural: EAMs must often decide not only which action is acceptable, but which data representation is ethically admissible for the present context.

Taken together, these lines show that context is not an auxiliary parameter. It is often the variable that determines whether adjudication should be rule-based, statistical, deliberative, or evidentiary, and whether the relevant object is an action, an output, a requirement, or a dataset.

6. Governance criteria, limitations, and open problems

Ethical adjudication is frequently mischaracterized as a technical substitute for governance. The auditing literature states the opposite. Ethics-based auditing is defined as “a structured process whereby an entity’s present or past behaviour is assessed for consistency with relevant principles or norms,” and the proposed criteria for successful procedures are holistic, traceable, accountable, strategic, dialectic, continuous, and driving re-design (Mokander et al., 2021). This framing places EAMs inside larger governance arrangements rather than above them.

Several recurring limitations follow from this. Declarative ethical programs make reasoning explicit, but the source paper identifies unresolved difficulties in formalizing fuzzy natural-language codes, choosing among multiple codes, representing multi-dimensional value rather than a single scalar utility, scaling to large domains, and combining learning with verification (Otterlo, 2017). REBAR, while quantitative, requires labor-intensive decomposition graphs, Key Factor tables, and log parsers; it also inherits simulation–reality gaps and recommends that symbolic constraints and verified rules remain primary even when LLMs are used for explanation (Diller et al., 18 May 2026). Content-moderation adjudicators may reward ethical refusals more than humans do, creating moderation bias, feedback loops, and normative lock-in if AI judges are used uncritically for evaluation and training (Pasch, 21 May 2025).

Benchmark-centered systems carry their own risks. EthicsMH is explicitly a pilot dataset, only 125 scenarios, synthetic in origin though professionally reviewed, culturally narrow, and “research-only, not diagnostic or prescriptive” (Kasu, 15 Sep 2025). ADEPT is transparent and replicable, but persona selection is value-laden, citations can be hallucinated, and panels can give an illusion of rigor while reflecting the underlying model’s priors (Zohny, 27 May 2025). Mirror improves consistency and professionalism in ethics review, yet authentic IRB data remain scarce because of privacy, and the authors recommend secure, institution-managed deployment with access controls and de-identified or synthetic training data (Ding et al., 9 Feb 2026).

These constraints support two broad conclusions. First, an EAM is best understood as a socio-technical component: it encodes norms, retrieves evidence, and structures decisions, but it also requires institutional choices about who sets the rules, how conflicts are escalated, what counts as sufficient evidence, and when human review is mandatory. Second, no single adjudication logic is adequate across domains. Majority aggregation, expected-utility optimization, rule execution, benchmark scoring, and multi-agent deliberation each capture different aspects of ethical decision-making. An EAM is therefore less a fixed algorithm than a governed interface between normative frameworks, evidentiary structures, and operational decisions.