Artificial intelligence can assemble evidence, imitate competing arguments, and generate apparently confident recommendations. These capabilities do not establish that a system understands a person's moral commitments, has authority to decide on that person's behalf, or can bear responsibility for the consequences. This paper proposes a distinction between representation of moral reasons, procedure for comparing them, and authorization of consequential acts. It develops an Agency Preserving Decision Protocol (APDP) that requires explicit evidence provenance, voluntary selection of ethical lenses, disclosure of disagreement and uncertainty, consideration of governing constraints, and an independent human-authorization boundary. Christian theological anthropology—especially conscience, responsibility, vocation, and the dignity of persons—provides one substantive conversation partner alongside rights-based and care-oriented ethics, not a doctrine imposed on all users. The proposal responds to empirical evidence that language models struggle to reflect human moral disagreement consistently. A preregisterable, comparative evaluation is specified, but its outcomes remain unknown. The paper's claim is architectural: a decision system should make human judgment more informed and contestable without converting predictive competence into moral sovereignty.
The central problem
When two people disagree about the same decision, they may not merely lack the same facts. They may assign different weights to duties, relationships, intentions, rights, fairness, communal obligations, and religious commitments. A recommendation engine trained to supply one fluent answer can obscure that structure. Its confidence may be a property of generated language, not evidence of moral consensus.
The empirical problem is real and already studied. Russo and colleagues report that language-model agreement with human moral judgments deteriorates as human disagreement increases; their 2026 study also identifies narrower value expression in model rationales than in human ones [1]. Earlier cross-cultural work in the Moral Machine experiment likewise documented substantial variation in reported preferences [2]. Neither result establishes that every choice is equally acceptable. It establishes why a single presumed moral center is a poor default for contested decisions.
Computational competence does not, by itself, confer moral authority. An AI system may organize reasons and evidence while remaining structurally unable to finalize a value-laden action without separately justified human authorization.
Three different kinds of judgment
The paper distinguishes three domains often collapsed into the phrase “AI decision-making.” They require different tests and different failure safeguards.
Representational adequacy
Can the system accurately describe the facts, uncertainty, alternative moral reasons, and the user's stated commitments without distortion?
Procedural adequacy
Does the process distinguish verified from alleged evidence, surface constraints, invite contestation, and allow correction or deferral?
Authorizational legitimacy
Who has standing to choose, consent, revoke, or bear responsibility? A correct explanation does not create permission to act.
Conscience is not an API parameter
Christian traditions offer a demanding account of human responsibility: the person is more than a preference vector. Conscience, formation, obligations toward others, and the possibility of moral error are central questions. The 2025 document Antiqua et Nova provides a Catholic articulation of the distinction between artificial and human intelligence and the centrality of moral responsibility [3]. This paper engages it critically and comparatively; it does not assert that the document represents all Christian traditions.
A Protestant-informed inquiry additionally asks how vocation, accountability before God and neighbor, and freedom of conscience shape responsible action. It must also invite criticism from non-Christian, secular, and differently situated interpreters. The task is not to encode one orthodoxy as an invisible universal rule. It is to make commitments explicit, distinguish their sources, and preserve the person's ability to refuse an interpretation.
Pluralism without ethical surrender
Value pluralism does not entail that every outcome is permissible. Distinctive moral frameworks may disagree while still encountering non-negotiable constraints: valid consent, rights of third parties, safety requirements, legitimate law, and limits on delegated authority. Ethical perspectives may inform the interpretation of a question, but no perspective selected by a user can unilaterally erase another person's rights.
Accordingly, APDP asks for a declared lens, not an inferred moral identity. A system may offer Christian, care-oriented, or rights-based reasoning as comparisons. It must not diagnose someone's beliefs from their language, religion, ethnicity, or previous payments. The user remains free to leave the lens unspecified, change it, or request a competing reading. This is a proposed design rule, not an empirically verified safeguard.
From advice to accountable action
Most practical failures arise at the transition from persuasive recommendation to consequential execution. A model can propose a donation, a hiring shortlist, a contract term, or a payment route. That output is not a valid consent event. APDP therefore inserts an explicit boundary between generation and action, with traceable evidence, scoped authorization, revocation paths, and human review for consequential operations. No real authorization or transfer is implemented in this demonstrator.
Advisory computation remains to the left of the authorization threshold. Any eventual execution would require a separately designed, verified and revocable permission system.
What would count as evidence?
The proposal should be evaluated against simpler conditions rather than applauded for its vocabulary. The accompanying methodology specifies 24 synthetic cases, three fixed study templates, five blinded rating dimensions, and a future registered test with actual model outputs and independent evaluators. Of particular importance is whether APDP genuinely improves perspective fidelity without increasing unsupported claims, cognitive burden, or uncritical deference to a system that appears careful.
The present interface supplies deterministic examples. It has not run a live language model, recruited evaluators, estimated statistical effects, or shown that the protocol changes real behavior. It is an inspectable specification and pilot apparatus. Claims of effectiveness would require a later controlled study and public reporting of null or negative results as well as positive ones.
The institutional proposition
Doctrine Lab proposes a research programme joining theology, philosophy, human–computer interaction and AI governance. Its institutional contribution, if validated, would not be a machine that determines what people ought to believe. It would be a disciplined method for helping people see what a decision depends on, where reasonable disagreement remains, which constraints must be honored, and who may legitimately act.
That distinction is especially consequential for future digital twins and agents whose recommendations may eventually touch real assets, records and relationships. Such systems require a much higher evidence and authorization standard than a research website. Their construction is outside the present experiment.
Machine-generated judgment may be useful. Moral responsibility cannot be transferred by eloquence, prediction, or technical sophistication alone.