A Contribution to the Debate Between Instrumentality and Substitution
Indice
1. Introduction: The Unresolved Antinomy
The spread of specifically Catholic AIs (Magisterium AI, Truthly AI) has generated a debate touching on fundamental questions of theological anthropology and ecclesiology. Abadal (2025) has formulated a critique rooted in Thomistic virtue theory and phenomenology of embodiment. Sanders (2025) has defended Catholic AI through comparative risk analysis. Umbrello (2025) has attempted synthesis through verifiable design requirements.
Despite the argumentative richness, the debate remains suspended on an antinomy: is AI a passive neutral tool (inadequate to mediate complex realities) or an active mediator (potential substitute for ecclesial authority)? This oscillation prevents both theoretical and pastoral resolution.
The present contribution maintains that the antinomy can be overcome through the concept of participated agency: AI is neither passive tool nor autonomous agent, but participates in human action of evangelization according to specific modalities that require rigorous theological discernment and architectural incorporation of ontological safeguards.
2. The Debate: Synthesis of Positions
2.1 Abadal: Four Fundamental Constraints
Abadal’s (2025) critique is articulated in four arguments with deep philosophical roots:
Structural epistemic fallibility. LLMs generate outputs based on statistical probability, not on truth. “Hallucinations” are intrinsically unavoidable. The claim of “accurate responses” is epistemologically problematic.
Distortion of authority. Automation bias leads to functional transfer of authority from Church to machine. In dense digital ecosystems, authority becomes a function of algorithmic social proof, shifting reliability from living tradition to technological platforms.
Impossibility of cultivating prudence. Recalling Aquinas (ST II-II, q. 47), Abadal (2025) emphasizes that prudence requires lived embodied experience. Machines do not live, do not suffer, cannot acquire or transmit prudence. Outsourcing discernment atrophies the capacities that catechesis aims to form.
Mystification through naming. Names like Magisterium AI generate categorical confusion. Catholic formation is an I-Thou reality (personal, sacramental), chatbots are I-It tools. Anthropomorphization habituates to seeking interpersonal goods from non-persons.
2.2 Sanders: Comparative Risk and Pastoral Urgency
Sanders (2025) contests through four counter-arguments:
The secular alternative is worse. Millions already consult LLMs for faith questions. The real choice is not “Magisterium AI vs priest” but “Magisterium AI vs secular/relativistic LLM.” Abandoning the field means ceding vital missionary territory.
Paradox of human fallibility. Abadal applies to AI standards of infallibility that no human catechist satisfies. Humans “hallucinate constantly – we call it error, malformed opinion.” Magisterium AI, trained on the Magisterium, demonstrates fidelity superior to many laity and ecclesiastics.
Crisis of scale. Faced with billions with existential questions and a global shortage of trained priests, the ideal of face-to-face encounter “is beautiful, but it is not the reality for the majority of the world.” We need “tireless digital missionaries, 24/7.”
Philosophy of the “off-ramp”. Magisterium AI is optimized as an “exit” toward tangible community (parish, Sacraments, Catechism), not for engagement. It is a “digital signal toward true encounter with Christ in the Church.”
2.3 Umbrello: Non-Substitutive Synthesis
Umbrello (2025) attempts synthesis through three moves:
Steelmanning Abadal. Translates the four constraints into permissibility requirements: containment (not elimination) of error, avoid habitual displacement, recognize impossibility of transmitting prudence, avoid ontological mystification.
Contextual differentiation. “Where AI is already the first stop, a non-substitutive Catholic AI can reduce harm compared to secular systems. Where AI is not the first stop, don’t elevate it; strengthen human channels.”
Technical specifications and governance. RAG with “no-source, no-statement,” verbatim citations, confidence labels, escalation protocols, session caps, public metrics (Source Coverage Rate, Refusal Rate, Off-Ramp Rate).
2.4 The Limit: Unresolved Oscillation
The debate does not resolve the fundamental tension. Abadal leans toward instrumental conception but criticizes automation bias (implicitly recognizing active mediation). Sanders pragmatically defends but does not ontologically articulate the nature of mediation. Umbrello proposes “non-substitutive” design but oscillates between AI as transparent informational tool and recognition of transformative capacity.
Open questions: Are hallucinations bugs or ontological limits? Does repeated use form virtuous or vicious habitus? Does AI “participate” in magisterial authority or “simulate” it? A category is needed that overcomes the tool/substitute binary.
3. Participated Agency: Ontological-Sociological Foundation
3.1 Beyond Dualism: Neither Passive Tool Nor Autonomous Agent
The concept of participated agency integrates Thomistic ontology of participation with the relational approach of STS.
From Thomism: asymmetric participation. Every created being participates in divine being according to specific modalities (ST I, q. 44, a. 1). God possesses being per essentiam, creatures per participationem. Analogically: God possesses infinite intelligence; human intellect participates through lumen naturale; AI participates in human intelligence in a derived-instrumental way.
Aquinas distinguishes principal cause and instrumental cause (ST I, q. 105, a. 5). The instrument possesses true causal efficacy but acts in virtue of the principal cause. AI operates as instrumental cause acting in virtue of human intelligence, but manifests second nature (ST I-II, q. 51): disposition acquired through training that generates operations not univocally determined by programming.
From STS: asymmetric co-constitution. Latour’s (2022) ANT shows that technologies are not neutral intermediaries but mediators that transform what they convey. AI actively co-constitutes possibilities of human action. But against radical symmetrization of ANT, we preserve ontological asymmetry at three levels:
Origin: AI systems are products of human intelligence
Teleology: They operate toward ends deriving from human intentionality
Responsibility: Only rational beings are capable of ethical action
3.2 Machine Habitus: Algorithmic Sedimentation
Machine habitus sociologically translates participated agency. It is not habitus of the machine (which does not possess a rational soul) but the process through which repeated interaction with algorithms forms stable dispositions in humans.
For Aquinas, habitus is a quality difficult to remove that disposes toward good or evil (ST I-II, q. 49, a. 2). Three elements: stability, moral orientation, acquisition through repetition.
Dual dynamic:
From humans to AI (sedimentation): ML algorithms incorporate biases, preferences, preexisting power structures from training data – crystallization of habitus malus in algorithmic form
From AI to humans (formation): Repeated use forms stable dispositions on where to seek information, how to formulate questions, which answers to consider authoritative
For Catholic AI, the distinction between:
Virtuous habitus: Familiarity with primary sources, habit of verification, disposition to seek deepening in ecclesial community
Vicious habitus: Dependence on immediate answers, atrophy of capacity to dwell in uncertainty, erosion of patience (cardinal virtue)
3.3 Three Criteria of Discernment
Criterion 1: Transparency of mediation. AI must constantly manifest its nature as instrumental cause: verbatim citations (not paraphrases), exposure of epistemic limits, explicit refusal of sacramental/personal categories.
Criterion 2: Systematic off-ramp. AI must operate as temporary bridge: each session ends with concrete ecclesial step, automatic escalation for sensitive categories, quantitative limits against dependency.
Criterion 3: Verifiable virtuous habitus formation. Metrics monitoring: usage frequency (detect dependency), complex/simple questions ratio (outsourcing prudence), conversion toward offline resources (bridge function).
4. Rereading the Debate Through Participated Agency
4.1 Abadal’s Constraints as Anti-Crystallization Criteria of Sin
Abadal’s constraints become criteria for identifying when participated agency degenerates into algorithmic crystallization of sin.
Constraint 1 → Ontological limit of derived participation. Hallucinations are not bugs but manifest the impossibility of participating in connaturalitas with revealed truth that requires divine illumination (Augustine, De vera religione XXXIX, 72). AI can process information about revealed truth but cannot participate in it in the proper sense.
Constraint 2 → Machine habitus and displacement. The risk is not consciously believing that Magisterium AI is the Magisterium, but developing habitual disposition to seek truth via conversational interfaces. This generates habitus malus: technocratic mode of relationship with faith that privileges immediacy over contemplation, information over transformation, propositional clarity over apophatic mystery.
Constraint 3 → Irreducibility of embodiment. AI can simulate patterns of prudential reasoning but lacks: (a) corporeality as pre-reflective schema (Merleau-Ponty 2003), (b) lived temporality as distention of the soul (Augustine, Confessions XI, 14-27), (c) affective connaturality with the good. Prudence requires incarnate presence to the particular that exceeds all formalization.
Constraint 4 → Ontological mystification. The name “Magisterium AI” obscures the participated (derived, instrumental, limited) nature of algorithmic agency. The conversational interface mimics I-Thou structure while operating as I-It, generating machine habitus of relationality that confuses technical mediation with personal encounter.
4.2 Sanders: Conditional Legitimacy
The comparative risk argument is valid in non-ideal conditions, but conditioned on AI operating as: (1) instrumental cause ordered to human finalities, (2) temporary bridge toward superior mediations, (3) formator of virtuous habitus.
Critique of the “more faithful” argument. Presupposes that fidelity reduces to propositional correctness. But in Thomism, fidelity to Tradition requires living connaturality, intellectual-affective habitus, existential witness. A catechist who occasionally errs but lives authentically is more faithful (in full theological sense) than an algorithm that cites perfectly but lacks spiritual life.
Qualification of “crisis of scale”. Presupposes catechesis reducible to scalable information transmission. But formation in faith requires personal mediations because it implies transformation of habitus through virtues acquired per connaturalitatem.
The “off-ramp” requires ontological incorporation. Not only intentional (designer wants) but architectonically incorporated: UX degradation for frequent users, non-bypassable hard caps, requirement of verifiable offline praxis.
4.3 Umbrello: Specification of Participated Agency
Umbrello’s (2025) specifications are architectural codifications of ontological asymmetry:
“No-source → no-statement” codifies: AI does not possess proper authority
Verbatim citations codify: AI mediates without appropriating
Confidence labels codify: epistemic limits of derived participation
“I Don’t Know” codifies: confession of ontological, not just epistemic limit
Escalation protocols recognize hierarchy of mediations: some realities require sacramental, personal incarnate mediations, charismatic connaturality – by ontological nature they exceed algorithmic mediation.
5. Design Principles for Authentic Participated Agency
5.1 Architecture: Codifying Ontological Asymmetry
Epistemological hierarchy of sources. Beyond standard RAG, implement tiers (Sacred Scripture/Symbols of faith → Definitive Magisterium → Ordinary Magisterium → Local Magisterium → Authoritative theology → Theological studies). Citations indicate tier. Contradictions between tiers: present both, emphasize hierarchy.
Ontological Boundary Recognition. Classifier identifying questions exceeding algorithmic mediation by ontological nature: sacramental, prudential, mystical, personal-pastoral categories. When query matches: immediate escalation with explanation that “this experience touches mystery that exceeds what algorithm can mediate.”
Habitus Formation Monitoring. System tracking individual usage patterns: frequency trend, complexity ratio, offline conversion, temporal pattern, return latency. Red flags activate interventions (cooldown, reflective messages, offline praxis requirements).
5.2 Session Design: Structural Off-Ramps
Every interaction must structurally (not rhetorically) orient toward exit. Virtuous pattern: response with citations → OFF-RAMP BLOCK with concrete next steps (read complete primary source, participate in formative event, deepen with human guide, reflect in silence) → session ends → next query available after cooldown or completion of verified offline action.
5.3 Youth Mode and Governance
Youth mode: Limited scope (only Tier 1-3), lowered OBR threshold, mandatory parental oversight, peer-oriented off-ramps, educational prompts on difference between information and life of faith.
Governance: Theological council (doctrinal oversight) → Ethical-technical board (design/metrics) → Operational team. Public quarterly reports with core metrics (Umbrello) + habitus metrics + case studies + red flags/corrective actions. Independent annual audit with binding recommendations.
Economic model: 100% free, 100% donation-based. No paywall (avoids exclusion of poor), no advertising (avoids commodification of attention). Donor incentivization aligned with mission: virtuous metrics (users directed to offline resources, reduction of technological dependence), not engagement metrics.
6. Conclusions
The debate on Catholic AI solicits fundamental theological reflection on the nature of intelligence, mediation, Christian formation in the algorithmic era. Abadal is right about real risks. Sanders is right about pastoral urgency in non-ideal conditions. Umbrello is right that the solution is rigorous design under verifiable constraints.
The framework of participated agency is not a compromise but a theoretical reconfiguration that overcomes the tool/substitute antinomy. Catholic AI is legitimate if and only if it operates as instrumental cause that: (1) participates in human intelligence without substituting it (asymmetric co-constitution), (2) authentically mediates toward permanent ecclesial relationships (ontological off-ramp), (3) forms virtuous dispositions that make the user progressively less dependent (habitus contra habitus malus).
This requires architectural incorporation of safeguards that codify: ontological asymmetry, temporariness of mediation, formation of virtuous habitus. Like every created reality, AI can mediate grace if ordered according to divine design, or crystallize sin if bent to logics contrary to the integral human. The distinction lies in the anthropological, ecclesiological, and teleological choices incorporated in algorithmic architectures.
A Catholic AI designed according to participated agency can function as humble ancilla theologiae digitale – servant, never master; bridge, never destination; participated instrument of a Church that always remains the living body of Christ entrusted to concrete persons who witness, love, accompany in flesh and blood.
References
Abadal, L.M. (2025) ‘Beware of Catholic AI’, Public Discourse, 11 November.
Augustine of Hippo (1997) Confessions. Turin: Einaudi.
Augustine of Hippo (2006) De vera religione, in Works of Saint Augustine. Rome: Città Nuova.
Dicastery for the Doctrine of the Faith and Dicastery for Culture and Education (2025) Antiqua et Nova, 14 January. Vatican City: Libreria Editrice Vaticana.
Latour, B. (2022) Reassembling the Social. Actor-Network Theory. Milan: Meltemi.
Merleau-Ponty, M. (2003) Phenomenology of Perception. Milan: Bompiani.
Sanders, J. (2025) ‘In Defense of Catholic AI’, response article.
Thomas Aquinas, Summa Theologiae (ST).
Umbrello, S. (2025) ‘Toward Non-Substitutive Catholic AI: Design Requirements and Governance’, working paper.