Generative AI in Ministry: Authorship, Formation, and Truthfulness in Christian Practice

Abstract

Generative language systems can now produce sermons, prayers, pastoral letters, Bible studies, and devotional material that congregations cannot reliably distinguish from human composition. This article asks what Christian ministry may and may not do with such tools, and argues that the decisive questions are about truthfulness, authorship, and formation rather than about capability. It begins with an accurate description of what these systems do - producing statistically plausible continuations of text without belief, understanding, or access to truth - and derives from that description the specific failure mode of fluent falsehood, which is particularly dangerous in a domain where hearers assume that citations of Scripture and tradition have been checked. It then distinguishes three categories of ministry activity: work that may be delegated to such tools without loss, work that may be assisted but not delegated because the preparation is itself formative for the minister, and work that cannot be delegated at all because its value consists in a person actually attending to another person. Prayer offered on someone's behalf, pastoral presence in grief, confession, and the preached word arising from a minister's own engagement with Scripture and congregation fall in the third category. The article addresses confidentiality, disclosure norms, the ethics of training data, and the theological anthropology that the technology makes newly urgent, and it proposes a policy framework congregations can adopt.

Research Question and Scope

Which activities in Christian ministry may legitimately be assisted or performed by generative language systems, which may not, and what theological criteria - concerning truthfulness, authorship, formation, and personal presence - determine where the boundary falls?

Method and Source Selection

The article begins with a technically accurate description of large language models sufficient for ethical evaluation, drawing on published research and technical documentation rather than on vendor claims or popular accounts. It deliberately avoids both inflated descriptions of machine understanding and dismissive characterizations that would misrepresent the systems' actual capabilities, since a sound ethical judgment depends on an accurate account of what is being judged.

Normative analysis proceeds from established Christian commitments concerning truthfulness, the nature of the pastoral office, the relation between a minister's own formation and their ministry, and the theology of prayer and presence. Scriptural texts are cited by book, chapter, and verse. Where Christian traditions differ - notably on the theology of ordination and on what may be delegated within the ministry of word and sacrament - the divergence is identified rather than resolved.

No system was tested, benchmarked, or evaluated for this article, no vendor or product is assessed, and no empirical claim is made about how frequently these tools are used in ministry or with what effect. The technology changes rapidly enough that capability-specific guidance would become obsolete quickly, so the framework is deliberately built on durable criteria rather than on the current state of any system.

1. What these systems do, described accurately

A large language model is trained to predict text. Given a sequence of tokens, it produces a probability distribution over what comes next, and text is generated by sampling from that distribution repeatedly. The training corpus is very large, the resulting models capture a great deal of structure about how language is used, and the outputs are frequently indistinguishable from competent human writing. Subsequent training stages shape the model's behaviour toward being helpful and toward following instructions, but the underlying operation remains the production of plausible continuations.

Two implications follow that are decisive for ministry use. First, plausibility and truth are different properties, and the system optimizes for the former. A statement that sounds like the kind of thing that would be true - a citation formatted correctly, a quotation attributed to a plausible author, a Greek word glossed confidently - is produced by the same process as a statement that is true, and the system has no internal mechanism that distinguishes them. This is why such systems generate confident falsehoods, commonly called hallucinations or confabulations.

Second, the system has no beliefs, intentions, experiences, or relationship to the reader. When it produces a prayer, it is not praying. When it produces a pastoral letter expressing sympathy, no one is sympathizing. When it produces a testimony, nothing happened to anyone. This is not a claim about whether machines could in principle have inner states; it is a description of what these systems are, and it holds regardless of how convincing the output is. The convincingness is precisely the point of the training.

It is equally important not to understate the capability, since a false picture in that direction produces bad decisions too. These systems summarize accurately much of the time, draft competently, translate usefully, restructure documents, generate alternatives, and answer many factual questions correctly. They are genuinely useful instruments. Dismissing them as merely autocomplete obscures the fact that they can perform a wide range of tasks well enough to be adopted, which is exactly why the ethical questions are pressing rather than theoretical.

The practical consequence of holding both descriptions together is that these tools are best understood as powerful and unreliable assistants whose output requires verification by someone competent to verify it. That formulation already excludes a large class of ministry uses, because it requires that a person with the relevant knowledge checks the output, and many proposed uses involve precisely the absence of such a person - the volunteer producing a Bible study on a passage they have not studied, or the minister generating material on a subject they do not know.

2. Fluent falsehood in a domain that assumes verification

The confabulation problem is unusually dangerous in religious use because of what hearers assume. When a congregation hears a preacher say that a Greek term carries a particular sense, that a church father wrote something, that a historical event occurred, or that a study found a result, they assume the claim has been checked. The claim's authority derives from that assumption. A generated text supplies the form of a checked claim without the checking, and the hearer has no way to detect the difference.

The specific failures observed with these systems are directly relevant. They produce citations to works that do not exist, attribute real quotations to the wrong authors, invent plausible-sounding statistics, misstate the content of biblical passages while citing them accurately by chapter and verse, and generate confident etymologies and semantic claims about Hebrew and Greek that are wrong. Each of these is precisely the kind of claim that carries weight in preaching and teaching and that most hearers are not equipped to check.

Scripture treats false speech about God with particular severity, and the relevant texts are not marginal. The prophets are condemned for speaking a vision of their own minds and saying thus says the Lord when he has not spoken (Jeremiah 23:16, 21; Ezekiel 13:6-7). Deuteronomy makes the presumptuous prophet's speech a capital matter (Deuteronomy 18:20). Paul refuses to falsify God's word (2 Corinthians 4:2). The category into which an unverified generated claim about Scripture falls is not innocent error; it is speaking with borrowed authority about something one has not established.

The obvious response - verify everything - is correct and is harder than it sounds, because verification is expensive precisely where the tool seemed most useful. Checking a generated claim about a Hebrew term requires the competence one was trying to substitute for. Checking a citation requires locating the source. In practice, the cost of verification often exceeds the cost of doing the work directly, and the temptation is to verify the claims one can check easily and let the others pass, which produces a text whose most impressive-sounding assertions are the least reliable.

A defensible working rule follows: generated content may be used where the user is competent to verify it and does, and may not be used where the user's reason for using it is that they lack the competence. This rule is uncomfortable because it withdraws the tool from exactly the situations where it looked most valuable - the volunteer with no theological training, the bivocational minister with no study time - but those are the situations in which unverified confident error enters congregational teaching, and the harm falls on people who trusted the source.

3. Authorship and the honesty a congregation is owed

When a minister preaches, the congregation reasonably believes they are hearing the fruit of that minister's engagement with the text, informed by that minister's knowledge of them. This belief is part of what gives the sermon its pastoral force: it is addressed speech from a known person who has been thinking about this text and about this congregation. A sermon substantially generated and delivered without disclosure trades on a belief the preacher knows to be false, which is deception whatever the quality of the content.

This is the same analysis that applies to sermon plagiarism, which the tradition has long regarded as a serious breach precisely because of the implied claim. The relevant test, as with plagiarism, is whether the congregation would feel deceived if they knew. A preacher who says openly that they used a tool to generate an outline, or to suggest illustrations, or to check the structure of an argument, has told the truth and the congregation can evaluate accordingly. A preacher who conceals it has made a judgment about what the congregation would think, and the concealment reveals the answer.

Disclosure norms should therefore be established by congregations rather than left to individual discretion, because individual discretion under time pressure reliably drifts. A workable standard distinguishes three levels: tools used in research and preparation in ways analogous to consulting a commentary need no disclosure; substantial drafting assistance should be disclosed as a general practice rather than sermon by sermon; and material presented as personal testimony, personal reflection, or pastoral address must not be generated at all.

The last category deserves emphasis because it is where the deception is most acute and the temptation most practical. A letter of condolence, a message to a member in crisis, a note to a family after a death, and a personal reflection in a newsletter all make an implicit claim: that the sender thought about this person. Generating such a message and sending it under one's own name asserts an attention that was not given. The efficiency gain is real and the thing exchanged for it is the substance of the pastoral relationship.

A related question concerns institutional rather than personal communication. Newsletters, policy documents, event descriptions, grant applications, and administrative correspondence carry no implicit claim about personal attention, and using generative tools for them raises no honesty problem provided the content is accurate and someone takes responsibility for it. Congregations should be careful, though, that the boundary does not erode: a pastoral letter from the minister included in a newsletter is a personal communication regardless of where it is printed.

4. Preparation as formation: what the minister loses

The strongest argument against delegating sermon preparation is not about the congregation but about the preacher. Wrestling with a text over a week - reading it repeatedly, noticing what is strange, consulting sources, finding that the obvious reading does not hold, being confronted by it personally - is the process by which a minister is formed by Scripture. The sermon is the visible output; the formation is the more important product, and it is precisely what is skipped when a draft is generated.

The tradition has consistently insisted on this ordering. Paul instructs Timothy to pay close attention to yourself and to your teaching (1 Timothy 4:16), putting the self first. Ezra is described as having set his heart to study the law of the Lord, and to do it, and to teach - the sequence is study, practice, then teaching (Ezra 7:10). Gregory the Great's Pastoral Rule devotes its first section to the character of the pastor before considering what should be said. The preacher who has not been addressed by the text is passing on information rather than testifying.

This gives a criterion that cuts more finely than a blanket prohibition. Uses that support the minister's own engagement - finding relevant secondary literature, checking a translation question, generating counter-arguments to test one's reading, summarizing a lengthy commentary to decide whether to read it - leave the formative work intact and may deepen it. Uses that substitute for the engagement - generating the exegetical conclusions, the structure, or the application - remove the very thing preaching is supposed to arise from.

The same principle applies beyond preaching. Preparing a Bible study forms the leader. Writing a pastoral letter requires thinking about the person. Composing a prayer for public worship requires attending to what this congregation needs to bring before God this week. In each case the labour is not merely a cost to be minimized; it is where the minister's own attention is directed toward God and toward people, which is a substantial part of what the pastoral vocation consists of.

Congregations have an interest here and can act on it. A minister who has no time to prepare will use whatever reduces the load, and the pressure that produces delegated preparation is usually structural rather than characterological: too many responsibilities, no protected study time, administrative work that has expanded to fill the week. A congregation that wants preaching arising from genuine engagement should provide the conditions for it, and should recognize that complaining about generated sermons while assigning the minister sixty hours of other work is not coherent.

5. Prayer, and what cannot be delegated

Generated prayer is where the category error is clearest. Prayer is address to God, and a system producing text in the form of a prayer is not addressing anyone. When such text is read aloud in worship by a person who means it, prayer is occurring - the person is praying, using words they did not compose, which is unremarkable since the church has always prayed using the psalms and written liturgies. When such text is produced and distributed as prayer without anyone praying it, nothing has occurred except the circulation of religious language.

The distinction matters most in intercession offered on someone's behalf. When a member asks for prayer, they are asking a person to bring their situation before God. A generated response, however well phrased, does not do this, and telling someone you have prayed for them when you generated text about them is a straightforward lie about a matter they care about. Paul's repeated assurances that he prays for his congregations - I do not cease to give thanks for you as I remember you in my prayers (Ephesians 1:16); we have not ceased praying for you (Colossians 1:9) - are claims about what he actually did.

Written liturgical resources occupy a different position and have a long history. Collects, litanies, and orders of service are composed by people, often centuries ago, and are prayed by congregations who did not write them. The composition is a legitimate craft, and there is no principled objection to a person using a tool while composing a prayer that they will then pray and that a congregation will pray. The objection arises when composition replaces rather than assists praying, and when the resulting text is presented as the offering of a person who never made it.

A related question concerns devotional material distributed to congregations - daily readings, reflections, meditations. These are presented as the fruit of someone's own reflection on Scripture, and readers use them to structure their own prayer. Generated devotional content presents a form without the substance, and it carries the confabulation risk into the private devotional life of members who will not check the claims. Congregations distributing such material should know how it was produced and should tell recipients.

Underlying all of this is a theological point about what prayer is. Prayer is not the production of appropriate words; it is a person turning toward God, which is why Scripture is unembarrassed about prayer that is inarticulate. The Spirit intercedes with sighs too deep for words, and the one who searches the heart knows what the mind of the Spirit is (Romans 8:26-27). Jesus warns against the assumption that prayer is improved by more words (Matthew 6:7-8). A technology that produces better religious language addresses a problem prayer does not have.

6. Pastoral presence and the simulation of attention

Pastoral care is the domain where delegation is most tempting and least defensible, because the demand is unlimited and the work is the attention itself. A minister cannot respond to every message promptly, remember every circumstance, or produce a thoughtful reply at three in the morning. Tools that draft replies, summarize a person's history, or suggest what to say address a genuine pressure. What they cannot do is the thing pastoral care consists of, which is one person actually attending to another.

The biblical picture of care is consistently embodied and specific. Weep with those who weep (Romans 12:15). Bear one another's burdens (Galatians 6:2). Job's friends sit with him on the ground seven days and seven nights, and no one spoke a word to him, for they saw that his suffering was very great (Job 2:13) - which is the one thing they do well before they begin talking. Jesus asks his disciples to stay awake with him in Gethsemane (Matthew 26:38-40), a request for presence rather than for words or solutions.

This suggests a clear boundary. Administrative support for pastoral work is legitimate: scheduling, tracking who has not been visited, summarizing one's own notes before a visit, drafting logistical communications. Substituting for the encounter is not: generating messages of sympathy, producing responses to disclosures of crisis, or maintaining an appearance of contact that no one is actually maintaining. The test is whether the person on the other end would feel their situation had been attended to if they knew how the message was produced.

Automated systems presented as pastoral resources - conversational agents offering spiritual guidance or crisis support - require particular caution. They may have genuine value in providing information, directing people to help, and offering something at hours when no person is available, and some people disclose more readily to a system than to a person. But they should be clearly identified as what they are, should not be presented as spiritual direction or counselling, and must have reliable escalation to human help. A person in crisis interacting with a system that cannot recognize the crisis is a foreseeable and serious harm.

Congregations should also consider what widespread use would do to expectations. If members come to expect immediate, articulate responses at any hour, ministers who respond as humans do - slowly, imperfectly, sometimes inadequately - will be found wanting. The technology can quietly reset the standard against which human care is measured, and the resulting pressure falls on ministers who are already the group reporting exhaustion. Setting explicit expectations about response times, and defending them, is a structural protection worth putting in place before the norm shifts.

7. Legitimate uses: where the tools genuinely help

A framework that only prohibits is neither honest nor useful, and there are substantial ministry uses that raise none of the concerns above. Administrative drafting - meeting agendas, event descriptions, policy documents, room booking correspondence, grant applications, job descriptions - carries no implicit claim about personal attention and requires only that someone accurate takes responsibility for the content. Congregations with no administrative staff can recover meaningful time here.

Summarization and search assistance are genuinely valuable where the user can verify. A minister deciding which of several commentaries to consult, working through a long denominational report, or locating discussions of a topic across their own past sermons is using the tool for triage rather than for conclusions. The verification burden is manageable because the output points to sources that can be read directly, and errors are caught by the subsequent reading.

Accessibility applications are among the strongest cases and deserve more attention than they receive. Automatic captioning of services for deaf and hard-of-hearing members, transcription of recorded teaching, simplified-language versions of documents for members with cognitive disabilities or limited literacy, and audio description all extend participation to people currently excluded. These uses serve the congregation's obligation to include, and the accuracy risk is manageable because errors are usually evident and correctable.

Translation deserves qualified endorsement. Machine translation can make announcements, pastoral information, and study materials available to members who do not share the congregation's dominant language, which is a real service to migrant and multilingual congregations. It should not be relied upon for Scripture, liturgy, doctrinal statements, or anything pastorally sensitive without competent human review, since translation errors in these domains carry consequences and the systems remain uneven across language pairs, particularly for languages with less training data.

Finally, these tools can serve preparation by supplying resistance. Asking for the strongest objections to a reading one has adopted, for alternative interpretations of a passage, or for how a text has been understood in traditions other than one's own can genuinely improve a sermon, precisely because it introduces material the preacher would not have generated. The output must be checked, but its function is to prompt further work rather than to conclude it, which keeps the formative labour intact.

Congregations with limited resources deserve a specific word here, because blanket caution can function as a privilege reserved for well-staffed churches. A bivocational minister serving a small congregation while working another job, a lay preacher with no theological training, and a church in a language with few available study resources are all in genuine difficulty, and telling them simply to do the work properly ignores their circumstances. The constructive response is not to relax the competence rule but to supply what is missing: denominations and networks can provide vetted study materials, accessible commentaries, and access to someone with training who will answer questions, which addresses the actual need rather than substituting a tool that produces confident error where competence is absent.

8. Confidentiality: what is typed into a prompt

Ministers considering these tools for pastoral work must understand where the text they enter goes. Depending on the service and its configuration, prompts may be transmitted to third-party infrastructure, retained, reviewed by humans for quality or safety purposes, and in some arrangements used to improve the systems. A minister who enters a member's situation in order to draft a response has disclosed that situation to parties the member never contemplated.

This is not a marginal concern given what pastoral work involves. Disclosures about abuse, addiction, mental illness, immigration status, sexuality, marital breakdown, and criminal matters are exactly the material a minister might want help responding to, and exactly the material that must not be transmitted. The relevant standard is the one applied to any other third party: information given in pastoral confidence is not shared without the person's knowledge and consent, and consent has not been given for this.

Even anonymization is insufficient in most realistic cases. Removing a name from a description of a specific pastoral situation does not remove the identifiability of the situation, and the combination of details that make it useful to describe is precisely the combination that identifies. Ministers who wish to use these tools for pastoral reflection should work with genuinely hypothetical situations rather than lightly disguised real ones.

Congregations should therefore establish a clear rule, and it should be simple enough to follow under pressure: no personal information about members, and no content of pastoral conversations, is entered into any external system. Where a congregation determines that some use is warranted - a service with contractual guarantees about retention and use, deployed within the organization's own controlled environment - that determination should be made deliberately by whoever is responsible for information governance, documented, and communicated, rather than made individually by whoever is drafting a reply.

Safeguarding material warrants an absolute prohibition. Records and disclosures relating to abuse or risk are subject to the strictest handling a congregation applies to anything, are frequently subject to legal requirements about storage and access, and may become evidence in proceedings. Processing them through an external system risks the integrity of the record, breaches the confidentiality owed to those involved, and may prejudice a subsequent investigation. This should be stated explicitly in congregational policy rather than assumed.

9. Provenance, labour, and the ethics of training data

These systems are trained on very large corpora of text produced by people, and the terms on which that material was collected are contested. Litigation and legislative activity across multiple jurisdictions concern whether the use of copyrighted works for training is lawful, whether authors and publishers are entitled to compensation, and what disclosure obligations apply. The legal position is unsettled and differs by jurisdiction, which means Christian users are making a moral judgment under genuine uncertainty rather than following a settled rule.

The relevant Christian commitments are clear even where their application is not. The labourer deserves to be paid (Luke 10:7; 1 Timothy 5:18), and Scripture treats the withholding of a worker's due with severity (Deuteronomy 24:14-15; James 5:4). Christian institutions that would not photocopy a book to avoid buying copies should think carefully before treating the aggregate use of authors' work as morally different simply because the mechanism is unfamiliar. At minimum, this is a matter for deliberate judgment rather than for adoption without reflection.

A second labour question concerns the workers who perform data annotation and content moderation for these systems, frequently in lower-income countries and often reviewing disturbing material under difficult conditions. Reporting on these practices has raised significant concerns about pay and psychological harm. Christian organizations that maintain ethical procurement standards for other goods have a consistent reason to ask the same questions here, and to prefer providers who can answer them.

Environmental cost is a third consideration that connects to Christian obligations regarding creation. Training and operating these systems consumes substantial energy and water, and the aggregate demand from data centres is significant enough to affect regional infrastructure. The magnitude is disputed and estimates vary widely, so the honest position is not that use is prohibited but that it is not free, and that an institution which speaks about creation care should count this among the costs it considers.

None of these considerations produces a straightforward prohibition, and pretending otherwise would be dishonest. Christians use many technologies with contested supply chains, and consistency would require far more than declining one tool. The reasonable conclusion is that congregations should make an informed decision rather than an unreflective one, should prefer providers with better practices where they can be identified, and should be willing to say publicly what they have decided and why.

10. Theological anthropology: what the technology makes newly urgent

These systems press a question the church has always answered but rarely had to defend: what is a human being, such that a machine producing indistinguishable text is nonetheless categorically different? The Christian answer does not rest on cognitive capability. It rests on the claim that human beings are created in the image of God (Genesis 1:26-27), known by God before formation (Psalm 139:13-16; Jeremiah 1:5), addressed by God, and capable of response. Dignity is conferred rather than achieved, which is why it does not diminish with capability.

This is a significant advantage of the Christian position and should be stated clearly, because the main secular alternatives ground human worth in rational capacity, autonomy, or contribution, and each of those grounds is threatened when machines exceed human performance. A framework in which worth derives from capability has no stable answer when capability is surpassed, and it has never had a good answer for humans whose capabilities are diminished by disability, illness, injury, or age. The imago Dei is not vulnerable in this way.

The corollary is that Christians should resist two symmetrical errors. The first is treating these systems as persons - attributing to them understanding, feeling, or moral status they do not have, and forming attachments to simulated relationship. The second is deriving from human uniqueness a licence for contempt toward the technology's users or a refusal to engage with what it changes. Neither error follows from the tradition, and both are common.

The idolatry question deserves careful statement rather than rhetorical use. The biblical polemic against idols turns on their being made by human hands and then treated as having power they do not possess: they have mouths but do not speak, eyes but do not see (Psalm 115:4-8; Isaiah 44:9-20). A system that produces speech without speaking, and that is treated by users as a source of wisdom, has a structural resemblance to this that is worth noting carefully. The resemblance is not an argument against use; it is a warning about the posture of the user.

Finally, the technology exposes something about how congregations have understood ministry. If a generated sermon is acceptable, then preaching was being valued as content delivery. If a generated pastoral message is acceptable, then care was being valued as communication received. The discomfort many feel at these substitutions is evidence of a better implicit theology than their practice sometimes reflects, and the appropriate response is to make that theology explicit rather than to argue about the tools.

11. Disclosure norms congregations can adopt

Congregations should adopt explicit disclosure norms rather than leaving individuals to judge, because the judgments are difficult, the pressure is real, and inconsistency between leaders will itself become a source of conflict. A workable norm has three tiers, corresponding to the analysis above, and can be stated in a paragraph that everyone in a ministry role can understand and apply.

The first tier is research and administrative use requiring no specific disclosure: consulting these tools as one would consult reference works, drafting administrative documents, and producing operational communications. The congregation should state as a general policy that such tools are used in administration, which is sufficient disclosure for material that carries no implicit claim of personal composition.

The second tier is substantial assistance with teaching and communication material, which should be disclosed as a standing practice. A congregation whose minister uses these tools in sermon preparation should say so once, publicly and clearly, describing what is and is not delegated. This protects the minister from later accusation, informs the congregation, and invites the accountability that makes the practice examinable rather than concealed.

The third tier is material that must not be generated: personal pastoral communication, testimony, prayer offered on behalf of a specific person, confession and its response, and anything presented as the minister's own reflection or experience. This should be an absolute rule rather than a presumption, because the harm is the deception itself and cannot be mitigated by quality. Congregations should state it plainly so that the expectation is shared rather than assumed.

Alongside disclosure, congregations need a competence rule and a confidentiality rule. The competence rule is that generated content on any matter of Scripture, doctrine, history, or fact is used only by someone able to verify it, and only after verification. The confidentiality rule is that no personal information about members and no content of pastoral conversations is entered into external systems. These three rules - disclosure, competence, confidentiality - cover the great majority of realistic situations.

12. Formation of congregations, not only policies

Policy governs institutional practice, but members are using these tools individually for devotional reading, Bible study, and questions about faith, and congregations that address only their own use have addressed the smaller part of the problem. A member asking a generative system what a passage means, or what the church teaches about a matter, will receive a fluent answer that may be accurate, may be subtly wrong, and will not indicate which.

The pastoral response is teaching rather than prohibition. Members can be taught what these systems do and do not do, shown examples of confident errors on biblical and theological questions, and given reliable alternatives - study Bibles, accessible commentaries, the congregation's own teaching, and access to someone who will answer questions. A congregation that supplies nothing has left members with the tool that is available, which is the situation that produces uncorrected error.

There is a formational opportunity here as well. The technology makes vivid a distinction the church has always wanted to teach: that knowing about God and knowing God are different, and that spiritual formation is not the acquisition of information. A system can produce accurate statements about prayer and cannot pray. It can summarize the doctrine of the Trinity and cannot worship. Teaching members why this distinction matters is teaching them something about their own discipleship, which is more useful than warnings about accuracy.

Congregations should also attend to members whose work is affected. Displacement, deskilling, and pressure to adopt tools they find ethically uncomfortable are live workplace realities for people in writing, translation, design, administration, customer service, and increasingly other fields. This is pastoral territory: people facing loss of vocational identity and material insecurity need care rather than commentary, and congregations that discuss the technology only as an ethical abstraction miss members who are living its consequences.

Finally, congregations should model the practice they teach. A church that instructs members about the limits of generated content while quietly using it for pastoral communication has taught the opposite of what it said, and the inconsistency will eventually be visible. The disclosure norms described above are as much a formational statement as a governance measure, since they demonstrate that the congregation regards attention to persons as something that cannot be simulated.

13. What ministry is for, and why the boundary falls where it does

The three-tier boundary proposed here - delegable, assistable, and non-delegable - is not derived from a judgment about capability, and this matters because capability will continue to improve. It is derived from what particular ministry activities are for. Administrative work exists to produce an accurate document, and a document produced with assistance is as good as one produced without. Preaching exists to bring a congregation before Scripture through a person formed by it. Pastoral care exists as attention given by one person to another.

This means the boundary does not move as systems improve, which is a substantial practical advantage over frameworks built on current limitations. A system that never confabulated would still not be praying when it produced a prayer, would still not have attended to a bereaved member when it produced a condolence, and would still not have subjected a preacher to the discipline of a difficult text. The formational and relational arguments are independent of accuracy.

The accuracy argument does move, and honesty requires acknowledging it. If confabulation were substantially solved, the competence rule would loosen for research and reference use, and some of the caution in this article would be excessive. Congregations should therefore distinguish the reasons for their policies, so that the parts contingent on current reliability can be revised as evidence warrants while the parts grounded in what ministry is remain stable.

A final observation concerns why this question feels urgent to ministers in a way that previous technologies did not. Printing, radio, and recording all extended the reach of ministry without touching its production; a printed sermon was still someone's sermon. These systems touch production, and they touch it at exactly the points where the labour is formative. The discomfort is therefore appropriate rather than reactionary, and it is pointing at something real.

The tradition's answer to what makes ministry ministry has never been the quality of its output. It has been that a person, called and formed, attends to God and to people on their behalf, and offers what they have actually received. Paul's account of his ministry is not that he produced excellent material but that he did not falsify God's word, did not use flattery, and shared not only the gospel but his own self (1 Thessalonians 2:8). What cannot be delegated is the self, and that is where the boundary falls.

14. Children, youth ministry, and the formation of the young

Young people are the heaviest users of these systems and the group whose formation is most affected by them, and congregational policy that addresses only ministerial practice omits them entirely. Children and adolescents use generative systems for schoolwork, for companionship, and increasingly for questions they would once have asked an adult. A young person asking such a system about faith, sexuality, doubt, or a family crisis receives an articulate answer from something with no responsibility toward them and no knowledge of their circumstances.

The formational concern is not primarily about misinformation, though that is real. It is about what is learned when every question receives an immediate, fluent, confident answer. Christian formation has traditionally involved sitting with unresolved questions, waiting, receiving instruction from people who know you, and discovering that some things are learned only over time. Deuteronomy's model of catechesis assumes a child's question answered by a parent within a relationship and a shared history (Deuteronomy 6:20-25). A system that answers instantly and anonymously teaches a different expectation about how understanding is acquired.

Youth ministry can respond constructively rather than defensively. Teaching young people to interrogate a generated answer - to ask where it came from, whether the citation exists, what a different tradition would say, and what the system cannot know about their situation - equips them for far more than religious questions. It is a form of the critical discernment that Paul commends when he tells the Thessalonians to test everything and hold fast to what is good (1 Thessalonians 5:21), and it treats young people as capable of judgment rather than requiring protection from information.

Congregations should also be candid with young people about companionship applications, which are marketed as relationship and are used by adolescents in significant numbers. The pastoral concern is not that they are dangerous in an obvious sense but that they offer the appearance of being known without any of the friction, cost, or reciprocity that actual relationship requires, and that habituation to this may make ordinary friendship seem unrewarding by comparison. This connects directly to the account of the body of Christ, whose members are precisely people one cannot mute or unfollow.

Finally, safeguarding policy should address these tools explicitly. Generated images can be used to produce abusive material involving identifiable children, conversational systems can be used to contact and groom, and young people may disclose to a system what they will not disclose to an adult. Congregations do not need technical expertise to respond sensibly: the existing principles - visibility, trained adults, escalation paths, and a culture in which young people know whom to tell - apply, and the policy should say so rather than leaving leaders to improvise.

Parents deserve support rather than alarm, and congregations are well placed to provide it. Many parents are aware that their children use these systems and have no framework for thinking about it beyond screen-time anxiety. A congregation that offers a plain explanation of what the tools do, some honest examples of their errors, suggested questions parents can ask their children, and the theological reason the church cares - that formation happens in relationships and takes time - has given families something more useful than a restriction they cannot enforce. This is ordinary catechetical work applied to a new occasion, and it treats parents as the primary formators that Deuteronomy 6 and Ephesians 6:4 assume them to be.

Limitations

• The capabilities and failure modes of generative language systems are changing rapidly. Descriptions here reflect general and well-documented characteristics rather than the current state of any particular system, and readers should expect specific claims about reliability to require revision. No system was tested or benchmarked, and no vendor or product is evaluated.
• The legal position on training data, copyright, and disclosure obligations is unsettled and differs substantially between jurisdictions, with active litigation and legislation in several. Nothing in this article constitutes legal advice, and the treatment of provenance identifies moral considerations under acknowledged uncertainty rather than stating settled conclusions.
• Christian traditions differ on the theology of ordination, on what may be delegated within the ministry of word and sacrament, and on the relation of the minister's person to the efficacy of their ministry. The framework here works from commitments the traditions largely share and will require adjustment in traditions with stronger or weaker accounts of the pastoral office.
• No empirical claim is made about how widely these tools are used in ministry, by whom, or with what effect on congregations, since no reliable data of that kind exists. Observations about pressures on ministers and about congregational expectations are drawn from the general literature on pastoral workload rather than from measurement of tool use.
• Environmental and labour considerations are stated as matters requiring deliberate judgment rather than resolved. Estimates of energy and water consumption vary widely by system, deployment, and method of accounting, and reporting on annotation labour concerns specific providers and periods rather than the field as a whole.

Conclusion

Generative language systems produce plausible text without belief, understanding, or access to truth, and this description rather than any judgment about their sophistication determines what ministry may do with them. The immediate danger is fluent falsehood in a domain where hearers assume that citations, translations, and historical claims have been checked, and the working rule that follows is that generated content may be used only by someone competent to verify it who does.

Three categories organize the rest. Administrative and operational work may be delegated, because it exists to produce an accurate document. Research and preparation may be assisted but not delegated, because the labour of engaging a text is how a minister is formed by it. Prayer offered for a person, pastoral presence, testimony, and personal address may not be delegated at all, because what they consist of is a person actually attending.

This boundary is stable as capability improves, which is its principal advantage. A system that never erred would still not be praying, would still not have attended to the bereaved, and would still not have subjected the preacher to the discipline of the text. Congregations should distinguish the parts of their policy that depend on current reliability, which may loosen, from the parts grounded in what ministry is, which will not.

The tradition's account of ministry has never rested on the quality of its output. Paul's summary is that he did not falsify God's word, refused flattery, and shared not only the gospel but his own self. What generative systems cannot supply is the self - the person who has been formed, who has attended, and who is present. Congregations that state this plainly, in disclosure norms they actually follow, will have said something true about themselves as well as about the technology.

References

1. The Holy Bible, New Revised Standard Version Updated Edition. (2021). National Council of Churches.
2. Gregory the Great. (c. 590). The Book of Pastoral Rule (Regula Pastoralis).
3. Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610-623.
4. Weidinger, L., et al. (2022). Taxonomy of risks posed by language models. Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, 214-229.
5. Ji, Z., et al. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), 1-38.
6. Floridi, L. (2023). The Ethics of Artificial Intelligence: Principles, Challenges, and Opportunities. Oxford University Press.
7. Herzfeld, N. (2023). The Artifice of Intelligence: Divine and Human Relationship in a Robotic Age. Fortress Press.
8. Graves, M. (2022). Apocalyptic AI and theological anthropology. Zygon: Journal of Religion and Science, 57(4).
9. Vanhoozer, K. J. (1998). Is There a Meaning in This Text? The Bible, the Reader, and the Morality of Literary Knowledge. Zondervan.
10. Smith, J. K. A. (2009). Desiring the Kingdom: Worship, Worldview, and Cultural Formation. Baker Academic.
11. Willard, D. (1988). The Spirit of the Disciplines: Understanding How God Changes Lives. Harper & Row.
12. Pontifical Academy for Life. (2020). Rome Call for AI Ethics.
13. Dicastery for the Doctrine of the Faith and Dicastery for Culture and Education. (2025). Antiqua et Nova: Note on the Relationship Between Artificial Intelligence and Human Intelligence.
14. Second Vatican Council. (1965). Gaudium et Spes: Pastoral Constitution on the Church in the Modern World.