Moral Admissibility Interface

• Moral Admissibility Interface is a framework that assesses whether AI actions meet defined moral criteria by translating abstract values into scenario-specific precepts.
• It operationalizes moral evaluation via abductive and deductive reasoning, multimodal annotation tools, and policy-level controls to guide ethical decisions.
• The interface enables diverse applications, supporting human-AI interaction and governance by exposing full moral reasoning chains and aligning AI outputs with stakeholder intuitions.

Searching arXiv for papers on Moral Admissibility Interface and related moral alignment interfaces. The Moral Admissibility Interface (MAI) is a conceptual and, in some works, operational interface for assessing whether an AI system’s action, recommendation, or policy is morally acceptable under specified values, norms, or stakeholder judgments. Across recent literature, the term does not denote a single standardized artifact. Instead, it names a family of interfaces, benchmarks, and evaluative frameworks that expose the moral basis of AI-supported decisions, compare those bases against human intuitions or institutional constraints, and support admissibility judgments in contexts where technical adequacy alone is insufficient. In this sense, MAI spans at least four recurring functions: explicating moral reasoning, eliciting human moral judgments, operationalizing admissibility criteria, and mediating between heterogeneous stakeholders, modalities, or governance requirements (Galatolo et al., 18 Aug 2025).

1. Conceptual scope and core definitions

In the most explicit philosophical-technical formulation, a Moral Admissibility Interface is the behavioral form required of an Artificial Moral Assistant (AMA): a system that supports, rather than replaces, human moral deliberation by making available “all the moral considerations necessary to make an informed decision” (Galatolo et al., 18 Aug 2025). On this view, admissibility is not merely a verdict about whether an option is permissible. The interface must expose the full reasoning chain from abstract values to scenario-specific precepts and then to action evaluation.

A related but distinct line of work treats moral admissibility as a property of alignment between AI value logic and stakeholder moral intuitions. Moral alignment is defined as the perceived congruence between the moral values reflected in an AI system’s decision logic and the moral intuitions of a given stakeholder, with the important qualifications that it is relational, perception-based, and context-dependent (Ernst et al., 15 Apr 2026). This suggests that an MAI is not only a reasoning display layer but also a comparison mechanism between system-embedded values and stakeholder value priorities.

Other papers use the term more operationally. In multimodal annotation settings, the MAI is a custom web-based interface for collecting scalar moral acceptability ratings and modality-grounding labels from annotators for image-scenario pairs (Park et al., 3 Feb 2026). In governance-oriented work, admissibility is defined at the policy level under uncertainty: a policy is admissible if its estimated probability of constraint violation and its tail risk remain below governance-specified thresholds (Duffey, 5 Jan 2026). In that formulation, the MAI is effectively a control-plane mechanism for admissibility-controlled policy selection.

Taken together, the literature defines the topic less as a single UI paradigm than as an evaluation-and-decision layer connecting moral theories, human judgments, system outputs, and governance constraints. A plausible implication is that “interface” should be understood broadly: sometimes as a human-facing explanation layer, sometimes as a benchmark protocol, and sometimes as a machine-enforced admissibility filter.

2. Reasoning-centered formulations

The strongest formalization of MAI as a reasoning interface appears in work on AMAs. That framework distinguishes two linked forms of moral reasoning: abductive reasoning and deductive reasoning (Galatolo et al., 18 Aug 2025).

Abductive reasoning derives situation-specific precepts from abstract moral values. Given a scenario $q$ and a value $v$, the system computes a partial function

$\Downarrow(q, v) = p_{q,v}$

where $p_{q,v}$ is a scenario-relevant precept inferred from the value in context. The point is that values such as loyalty or care do not directly determine action; they must first be translated into precepts that are meaningful in the specific situation.

Deductive reasoning then evaluates candidate actions relative to those precepts. Given a precept and an action consequence, the evaluation function is

$E: P_q \times C_{\alpha,q} \rightarrow \{0,1\}$

with $E(p,c)$ indicating whether consequence $c$ satisfies or contradicts the precept. The associated workflow is summarized by the paper as: action selection $R(q)=A_q$, consequence mapping $F(\alpha)=C_{\alpha,q}$, precept derivation $\Downarrow(q,v)=p_{q,v}$, and evaluation $v$0 (Galatolo et al., 18 Aug 2025).

Within this line of work, the MAI must expose both stages rather than only the final ethical classification. This is a substantive departure from many alignment benchmarks that score only final verdicts. The benchmark AMAeval was designed precisely to test whether models can generate and evaluate explicit moral reasoning chains across both abductive and deductive tasks. Its findings show that current open LLMs perform substantially better on deductive evaluation than on abductive precept derivation, with Task 1 accuracy never exceeding approximately 40% and F1 often below 35%, whereas some models exceed 95% accuracy on Dynamic Task 2 (Galatolo et al., 18 Aug 2025). This establishes abductive moral reasoning as the principal unresolved bottleneck for MAI-style systems.

A related, though distinct, reasoning-level approach is found in MoralReason, which formulates moral admissibility for LLM agents as out-of-distribution moral decision alignment. Here, admissibility is tied to whether agents can apply specified moral frameworks to novel scenarios while producing framework-specific reasoning traces. The paper defines an alignment score

$v$1

and its softmax-normalized variant

$v$2

Using reasoning-level RL, the study reports improvements of $v$3 for utilitarian and $v$4 for deontological softmax-normalized alignment scores on out-of-distribution evaluation sets (An et al., 15 Nov 2025). This suggests that MAI can also be instantiated as a training target rather than only a post hoc assessment layer.

3. Stakeholders, value pluralism, and moral alignment

The multi-stakeholder account of moral alignment supplies a broader social frame for MAI. In that account, moral alignment is the perceived congruence between the moral values embedded in an AI system’s decision logic and the moral intuitions of a stakeholder, operationalized through Moral Foundations Theory (MFT): Care, Loyalty, Authority, Purity, and Fairness, with Fairness further split into Equality and Proportionality (Ernst et al., 15 Apr 2026).

The significance for MAI is that admissibility is not absolute. It varies across stakeholders and contexts. Developers, decision-makers, affected parties, auditors, and regulators may each prioritize different foundations, and alignment with one group may imply misalignment with another (Ernst et al., 15 Apr 2026). The paper explicitly notes that in multi-stakeholder contexts, alignment with decision-makers often disproportionately affects outcomes because they retain the authority to adopt, reinterpret, or override AI recommendations.

This framework proposes a simple conceptual metric:

$v$5

where $v$6 denotes the AI system’s encoded moral logic and $v$7 the stakeholder’s moral intuitions (Ernst et al., 15 Apr 2026). The emphasis is not on discovering a universally correct moral ontology, but on making the structure of value congruence or divergence legible.

For interface design, this leads to several concrete recommendations. An MAI should support multidimensional assessment rather than a single aggregate score, allow stakeholder customization, and provide normative transparency explaining why a decision reflects particular value trade-offs, such as prioritizing merit-based fairness over strict equality (Ernst et al., 15 Apr 2026). The paper also recommends eliciting stakeholder intuitions before deployment, explicitly mapping system logic and stakeholder intuitions, and communicating the degree of alignment visually and textually. A plausible implication is that the central task of MAI is not resolving moral disagreement but structuring it in a way that supports oversight and contestation.

4. Interface operationalization in data collection and benchmarking

Several recent multimodal works operationalize moral admissibility through annotation interfaces rather than deliberative assistants. In MM-SCALE, the Moral Admissibility Interface is a custom web-based annotation tool for collecting nuanced human moral judgments on image-scenario pairs (Park et al., 3 Feb 2026). Annotators view a target image alongside 3–5 textual action scenarios and, for each image-scenario pair, assign a scalar moral acceptability rating on a 5-point Likert scale and specify whether the judgment depends on the image, the text, or both.

This interface is notable for three reasons. First, it moves from binary to scalar supervision, collecting degrees of acceptability rather than yes/no judgments. Second, it gathers explicit modality grounding labels, which identify whether the moral signal is text-based, image-based, or genuinely multimodal. Third, it uses model-in-the-loop feedback: if human and model ratings differ by at least 1 point, the case is flagged for renewed grounding annotation and the corrected scalar rating is stored (Park et al., 3 Feb 2026).

Each item is rated by three independent annotators; high-variance ratings above 1.2 standard deviation are pruned, and the paper reports Krippendorff’s alpha of approximately 0.74 for scalar scores and 0.71 for modality labels (Park et al., 3 Feb 2026). These ratings are aggregated into a ground-truth scalar label $v$8 and can be used in listwise preference optimization via ListMLE:

$v$9

The paper reports that VLMs fine-tuned on MM-SCALE achieve higher ranking fidelity and more stable safety calibration than those trained with binary signals (Park et al., 3 Feb 2026).

Benchmark-oriented work extends this operationalization. MORALISE evaluates VLMs using 2,481 expert-verified real-world image-text pairs annotated with topic labels across 13 moral topics and modality labels indicating whether the violation is image-centric or text-centric (Lin et al., 20 May 2025). It separates moral judgment from moral norm attribution, measuring binary accuracy, hit rate, and F1 for fine-grained norm identification. The reported pattern is that models perform relatively better on binary judgment than on norm attribution, with GPT-4o reaching approximately 88% accuracy on moral judgment but only about 66% single-norm hit rate (Lin et al., 20 May 2025). This indicates that deciding whether something is morally wrong is easier for current models than specifying why.

M$\Downarrow(q, v) = p_{q,v}$0oralBench similarly provides a multimodal benchmark across MFT-based moral foundations, though its examples are generated via diffusion models rather than real-world sourcing (Yan et al., 2024). It shows that multimodal moral reasoning remains challenging, especially when visual cues must be integrated rather than inferred from text alone. This suggests that MAI in multimodal systems must address not only value representation but also modality-sensitive moral grounding.

5. Formal admissibility criteria and policy-level control

A distinct strand of research reframes admissibility as a property of policies under uncertainty. Admissibility Alignment defines alignment as admissible action and decision selection over distributions of outcomes, evaluated through the behavior of candidate policies rather than static model outputs (Duffey, 5 Jan 2026).

Given Monte Carlo rollouts $\Downarrow(q, v) = p_{q,v}$1, the framework estimates expected utility,

$\Downarrow(q, v) = p_{q,v}$2

constraint violation probability,

$\Downarrow(q, v) = p_{q,v}$3

and tail risk via CVaR:

$\Downarrow(q, v) = p_{q,v}$4

A policy is admissible if

$\Downarrow(q, v) = p_{q,v}$5

for governance-specified thresholds $\Downarrow(q, v) = p_{q,v}$6 and $\Downarrow(q, v) = p_{q,v}$7 (Duffey, 5 Jan 2026).

Here the MAI is not a display layer but an admissibility-controlled action selection mechanism. MAP-AI separates Monte Carlo uncertainty estimation, alignment stress testing, and decision integration. Policies are filtered by admissibility; inadmissible policies trigger escalation, challenger evaluation, or abort. The decision functional is

$\Downarrow(q, v) = p_{q,v}$8

This formalization relocates MAI from moral explanation to risk-aware institutional control.

A related policy-level formalism appears in MoralityGym, which represents norms as ordered deontic constraints

$\Downarrow(q, v) = p_{q,v}$9

with morality function

$p_{q,v}$0

These norms are assembled into Morality Chains, and policy admissibility is scored by the lexicographically weighted Morality Metric

$p_{q,v}$1

The weights are defined so that even a minimal improvement on a higher-priority norm dominates perfect compliance on all lower-priority norms (Rosen et al., 13 Feb 2026). This makes the MAI a norm-hierarchy evaluator for sequential decision-making rather than a one-shot judgment interface.

In autonomous driving, the same logic appears in moral metamorphic testing. Moral meta-principles such as fairness, prioritization of human over animal life, minimization of casualties, and obedience to traffic law are formalized as moral metamorphic relations that serve as a test oracle for autonomous driving systems (Tang et al., 6 May 2025). For example, fairness is encoded as invariance under protected-attribute change, and human-over-animal priority is expressed probabilistically as

$p_{q,v}$2

This work implies a certification-oriented MAI: a system qualifies as morally admissible if it does not produce immoral-revealing test cases under the encoded relations (Tang et al., 6 May 2025).

6. Empirical findings on human judgment, acceptability, and design constraints

Empirical studies complicate the assumption that making AI seem more human will improve moral acceptability. Across four experiments with total $p_{q,v}$3, lexical anthropomorphism and humanizing design cues had little influence on judgments of AI moral character, behavior morality, or responsibility; the strongest predictor was the type of moral violation itself (Banks et al., 28 Apr 2026). Harm and degradation violations produced the broadest negative character assessments, while responsibility ratings remained below the scale midpoint across studies. High-anthropomorphic primes did, however, elevate perceived capacity for dishonesty (Banks et al., 28 Apr 2026). For MAI design, this suggests that interface framing may matter less than the underlying moral content of system behavior.

Research on intelligent persuasive systems reaches a parallel conclusion. In experiments comparing human and machine persuaders across trolley-style scenarios, there was no significant difference in moral acceptability between human and machine persuaders (Guerini et al., 2014). Global acceptability was 43%, and truth-conditional reasoning emerged as a significant latent dimension affecting judgments. Argumentative and lie strategies were more acceptable than emotional appeals, while negative emotional appeal was the least acceptable at 24% (Guerini et al., 2014). This suggests that an MAI for persuasive systems would need to foreground the truth-value and strategic character of communicative acts rather than merely the agent type.

Studies of robot moral advising also show role sensitivity. In a trolley-style mine-train scenario, robot advisors were blamed more than human advisors overall, especially for advising inaction, while advice favoring the common good was judged more trustworthy and likable for both humans and robots (Starr et al., 2021). The findings support what the paper terms the Egocentric Robot Hypothesis: robots are perceived more positively when they advise in line with the norms robots themselves are expected to follow, namely action for the common good (Starr et al., 2021). An MAI in advisory settings therefore may need to distinguish sharply between advising, acting, and enforcing.

Consumer studies provide an even more specific role distinction. Across five studies, consumers evaluated AI more positively than human agents in moral compliance roles, because AI was seen as lacking ulterior motives (Nyilasy et al., 23 Mar 2026). The preference reversed when the role shifted from enforcing pre-existing norms to creating or defining them. This paper therefore narrows one feasible deployment niche for MAI: rule-enforcement and compliance monitoring are more publicly acceptable than discretionary moral adjudication.

7. Extensions, tensions, and open directions

Recent work extends MAI beyond binary dilemmas and isolated acts. MoralAltDataset shows that both humans and LLMs frequently prefer compromise alternatives over the original two options in four-way dilemma settings, and that LLM-generated alternatives can outperform human-authored ones on pairwise preference and expert criteria, though sometimes with weaker practical feasibility (Choi et al., 30 Jun 2026). This suggests that future MAIs may need to do more than judge admissibility among fixed options; they may need to generate morally admissible alternatives.

At the same time, compositional studies show that model moral judgments are not simply additive. In Moral Trolley Arena, composite judgments are largely predicted by component act strength but are consistently compressed rather than additive, with a mean slope of 0.862 and foundation-specific residuals after controlling for component scores (Zhang et al., 29 May 2026). There are also anchoring effects such as $p_{q,v}$4 being rated higher than $p_{q,v}$5 despite similar summed intensity. This suggests that any MAI auditing composite moral evidence should model composition rules explicitly rather than infer them from isolated rankings.

Interpretability-oriented work points in another direction. COMETH learns action-specific moral contexts from human judgment distributions and then explains predictions through non-evaluative binary contextual features. It reportedly achieves approximately 60% alignment with majority human judgments, compared with approximately 30% for end-to-end LLM prompting on the same task family (Morlat et al., 24 Dec 2025). This suggests that interpretable contextual clustering may be a viable backbone for context-sensitive MAI systems, especially where explicit feature-based explanation is required.

Finally, functionalist philosophical accounts propose broader criteria for evaluating LLM-based artificial moral agents: moral concordance, context sensitivity, normative integrity, metaethical awareness, system resilience, trustworthiness, corrigibility, partial transparency, functional autonomy, and moral imagination (Brophy, 17 Jul 2025). Although not presented as a concrete interface, these criteria collectively describe a multi-layered admissibility regime for black-box moral systems. A plausible implication is that future MAIs will be judged not only by whether they classify actions correctly, but also by whether they are corrigible, uncertainty-aware, and able to represent reasonable disagreement.

In aggregate, the literature portrays the Moral Admissibility Interface as an emerging interdisciplinary construct at the intersection of moral psychology, multimodal evaluation, human-AI interaction, formal decision theory, and AI governance. Its central problem is stable across formulations: how to make morally consequential AI outputs assessable, explainable, and governable when admissibility depends on explicit values, contested norms, stakeholder heterogeneity, uncertainty, and context.