Who Is the Witness? Artificial Intelligence, Derivative Witness, and the Theology of Christian Proclamation

1. Introduction

Artificial intelligence has moved from administrative support into sermon preparation, doctrinal explanation, and pastoral communication. Reyes’s 2026 field experiment with pastors found that AI-assisted sermon drafts could be perceived as clear and spiritually encouraging when pastors actively shaped and contextualized them, while weaknesses remained in exegesis, doctrinal precision, contextual illustration, and pastoral ownership (Reyes 2026). Mannerfelt and Roitto’s (2025) study of Swedish priests reaches a related conclusion by a different route: none of the preachers they observed accepted AI output uncritically, but their preachers often lacked a clear model of what the technology could and could not do, leading them to both underestimate and overestimate its communicative capacities. Taken together, these studies establish an empirical baseline this article does not need to relitigate. The practical question is no longer whether AI can produce sermon-like language. It can, and pastors are already using it.

Epistemology is simultaneously debating artificial testimony. He and Yang argue that LLM-generated statements can function as a source of testimonial knowledge under conditions suited to artificial rather than human testifiers (He and Yang 2025). That debate concerns whether a recipient can justifiably learn from AI output. Christian proclamation adds another issue, one that epistemology is not equipped to settle on its own terms: whether the source performs an act of witness at all, independent of whether a hearer can be justified in believing what the source says.

This article distinguishes proclamation-content from proclaiming agency. A text may contain true Christian claims while lacking a witness who believes, confesses, promises, repents, loves, or accepts responsibility for those claims. The proposed category of Derivative Witness explains how AI-generated words can nevertheless become part of a human act of proclamation through verification, assent, appropriation, and answerability. The argument proceeds in six stages. It first separates artificial testimony from Christian witness on independent philosophical and theological grounds. It then develops Derivative Witness as a positive category, specifying what each of its four conditions requires and what happens when they are only partially met. It applies the category to six cases spanning preaching and prayer, from the clearest failure to the clearest success. It then works through objections a sympathetic and a skeptical reader would both raise. It draws out practical-theological consequences for disclosure, formation, and pastoral training. It closes by naming the limits of the proposal and the empirical and theological work that remains.

2. Testimony and Artificial Output

He and Yang’s Reidian account deliberately loosens anthropocentric assumptions about testimony. They argue that artificial output can justify belief when a system robustly manifests artificial analogues of veracity and cautiousness (He and Yang 2025). This is an epistemic proposal about knowledge transmission. It does not require the machine to possess the full interpersonal psychology of a human witness, and the authors do not claim that it does. Their question is whether a hearer can be justified in believing what a system says. The present article’s question is different and, on their own terms, orthogonal to it: whether the system, in producing that output, performs an act of Christian witness.

That distinction is exactly why theology should not simply import “artificial testimony” as “Christian witness.” Christian witness ordinarily includes more than reliable content delivery. To witness to Christ is not merely to emit a proposition about Christ; it is an act located within discipleship, confession, communal accountability, and embodied agency. A recording can preserve a witness’s words without becoming the witness. A translation can carry testimony without becoming the testifier. Neither observation is original to this article. What follows is an attempt to say precisely why the observation holds, rather than simply asserting it by analogy.

2.1. The illocutionary requirement

Speech-act theory supplies the sharpest available tool for this purpose. Austin distinguishes the locutionary act of producing a meaningful utterance from the illocutionary act performed in uttering it, such as promising, asserting, or confessing, and from the perlocutionary act of bringing about an effect in a hearer, such as persuading or comforting (Austin 1962). Searle systematized this distinction and argued that illocutionary acts are individuated by their constitutive rules, including a sincerity condition: to assert that p is, among other things, to represent oneself as believing that p; to promise is to represent oneself as intending to keep the promise (Searle 1969). An utterance can satisfy the locutionary conditions of an illocutionary act, meaning it has the right grammatical and semantic form, while failing the act’s sincerity condition. A person who says “I promise” without any intention of keeping the promise has still performed the locutionary act of uttering the words, and has arguably performed a defective illocutionary act of promising, but has not thereby placed themselves under the same moral standing as a sincere promisor. The infelicity is not cosmetic.

Christian witness, on the account developed here, is best modeled as an illocutionary act with a demanding sincerity condition. To testify that Christ is risen is, among other things, to represent oneself as believing that Christ is risen, as having some standing to say so, and as accepting whatever consequences follow from having said it in that setting. An AI system that outputs the sentence “Christ is risen” satisfies the locutionary requirement. It produces a well-formed, semantically coherent utterance with the correct grammatical form of an assertion. It cannot satisfy the sincerity condition, because there is no representable belief state for the system to be representing. This is not a claim about the system’s processing being too simple or its outputs too formulaic; more sophisticated systems would not close this gap, because the gap is not one of sophistication. It is a claim about what the sincerity condition of the illocutionary act of witnessing requires, namely a believing, confessing subject, and about what current systems are not evidenced to have. Present systems generate language through computational processes without established evidence of religious belief, commitment, worship, repentance, or moral answerability. Theological agency should therefore not be inferred from linguistic performance alone, however fluent.

2.2. The counter-tradition: speech through unlikely sources

A fair hearing of the opposing intuition requires engaging a real counter-tradition rather than dismissing it. Scripture itself contains cases of proclamation-content issuing from a source that, on the face of it, lacks the ordinary psychology of a witness. Balaam’s donkey speaks true words of rebuke without being a moral agent capable of belief in the theologically loaded sense (Numbers 22:28-30). Caiaphas prophesies that one man should die for the people, and the Gospel of John explicitly attributes this to his office rather than to his personal conviction, since he spoke it “not of himself” (John 11:49-51). These cases matter because they show the tradition already has categories for true, even providentially significant, speech that does not originate in the speaker’s own settled belief.

The disanalogy with an AI system is nonetheless real and worth stating precisely rather than assuming. Balaam’s donkey and Caiaphas are, on the traditional reading, instances of God overriding or working through a creature’s speech for a specific providential purpose, disclosed within a narrative that identifies the speech as extraordinary and marks it as such for the reader. Neither case is offered by Scripture as a model for ordinary proclamation, and neither is proposed by any tradition this article is aware of as a general method by which the church should expect to receive its everyday preaching. An AI system’s sermon output is not narratively marked as a special providential override; it is presented, when it is presented badly, as ordinary preaching indistinguishable in form from a human sermon. The theological tradition’s willingness to accept rare, narratively signaled exceptions to the rule that witness requires a believing subject is therefore not evidence that the rule should be abandoned for a technology whose outputs are, by design, meant to blend into ordinary use. If anything, the existence of these two carefully marked exceptions in the canonical text argues for the rule’s ordinary strength: exceptional cases are exceptional precisely because the default requires a witness who believes.

3. Derivative Witness

Derivative Witness occurs when generated language becomes incorporated into the responsible communicative act of a human witness. Four conditions are required. Verification: the human checks the theological, factual, and pastoral content. Assent: the human actually affirms what will be said. Appropriation: the human intentionally makes the language part of his or her own proclamation rather than merely relaying it blindly. Answerability: the human accepts responsibility for the effects, errors, promises, and claims made in public.

The four conditions explain why identical sentences can have different theological status in different settings. A pastor who reads an unknown AI-generated sermon without checking it lacks verification and perhaps assent. A pastor who uses an AI draft as raw material, checks every claim, rewrites it within the congregation’s situation, and publicly owns the resulting sermon may use generated wording within a genuinely human act of proclamation. The witness is the accountable preacher, not the software.

The term “derivative” does not imply inferiority. Human proclamation has always used derivative materials: Scripture translations, commentaries, inherited liturgies, hymns, quotations, editors, and collaborative preparation. The distinct problem with AI is opacity about agency. Because its language can mimic first-person conviction, hearers may mistakenly attribute commitment to a system that has none.

3.1. Are the four conditions independent?

Verification and assent can seem close enough to collapse into one condition: checking content and affirming it look, on a quick reading, like two descriptions of the same act. They come apart in a case worth stating explicitly. A pastor may verify that an AI-generated illustration is factually accurate, that its Scripture citations are correctly rendered, and that no doctrinal claim in it is heterodox, and may nonetheless decline to assent to using it, on the ground that its tone does not fit the occasion or that a true and orthodox illustration is still the wrong illustration for this congregation on this Sunday. Verification is a check on the content’s adequacy by external standards, truth, accuracy, doctrinal fit. Assent is the pastor’s own act of affirmation, which can withhold itself even from content that has cleared every external check. The two conditions are therefore genuinely distinct, and a design that treats them as one risks collapsing exactly the space in which a pastor’s own judgment, as opposed to a fact-check, operates.

Appropriation is defined here by contrast with blind relaying, but a positive account is owed. Appropriation occurs when the language, once adopted, functions in the sermon as an expression of the preacher’s own conviction rather than as a quoted or bracketed external source. A preacher who says “an AI assistant suggested the following illustration, which I want to share with you” has disclosed the source without necessarily appropriating the content; the illustration remains, in the hearers’ understanding, something introduced from outside rather than spoken in the preacher’s own voice. A preacher who takes the same illustration, reworks its details to fit a parishioner’s actual situation, and delivers it without attribution as an example drawn from pastoral experience has appropriated it, for better or worse, whether or not this is disclosed elsewhere. Appropriation is therefore partly a matter of rhetorical form, how the content is framed for the hearer, and partly a matter of the preacher’s own relation to it, whether the preacher would defend the content as their own considered word if challenged afterward.

Answerability is the condition most resistant to third-party verification, since accepting responsibility is an internal disposition rather than an observable act. Two markers make it more tractable without making it fully checkable. First, a preacher who has genuinely accepted answerability will correct an error publicly and promptly when one is identified, rather than attributing the error to the tool that produced it. Second, a preacher who has genuinely accepted answerability will be willing, in principle, to explain and defend the claim in a setting where defending it carries cost, such as a pastoral conversation with someone the claim has hurt. Neither marker settles every case in advance, and both are retrospective: they can confirm that answerability was present after the fact, in a correction or a difficult conversation, more easily than they can certify it in advance of any test.

3.2. Partial satisfaction and degrees of derivative witness

The four conditions are presented in Section 3 as a set that is either met or not met for a given sermon. Real cases are frequently partial. A pastor may verify content thoroughly, assent to most of it, appropriate the majority of the sermon, and accept answerability for the whole, while one paragraph retains the AI’s original phrasing essentially unedited because the pastor judged it accurate and well put. This article does not propose a threshold, such as a percentage of original wording, above which Derivative Witness is achieved and below which it fails. The four conditions function better as a diagnostic grid than as a pass-fail gate. A sermon can be strongly derivative-witnessed on three conditions and weakly so on the fourth, and the honest theological description of such a sermon is exactly that: strong on three, weak on one, rather than a single verdict that erases the difference.

This has a consequence for how the category should be used pastorally. A denomination or oversight body applying the four conditions as a compliance checklist, satisfied or not satisfied, will misuse the category. The conditions are better used as a set of questions a preacher asks of their own preparation process, and that a mentor or peer reviewer might ask in a formation setting, than as an audit instrument applied after the fact to a finished text. A finished sermon manuscript does not show its own history of verification, assent, appropriation, and answerability; those are properties of the process that produced the text and the disposition of the person who delivered it, not properties readable off the text itself.

3.3. Collective and institutional derivative witness

The framework as stated addresses a single human agent. Much real preaching involves more than one person: a preaching team drafts collaboratively, an associate pastor prepares a manuscript that a senior pastor delivers, a denomination supplies approved liturgical language that a local minister reads. Distributed authorship of this kind long predates AI and is not, by itself, a threat to genuine witness; a senior pastor who delivers a sermon drafted by a trusted associate can still verify, assent to, appropriate, and answer for it, exactly as a pastor can do these things with an AI draft. What changes when a team is involved is that the four conditions must be satisfied by whoever stands before the congregation and speaks in the first person, not merely by whoever did the drafting. A drafting team’s own verification of a sermon’s content does not discharge the delivering preacher’s obligation to verify it again for their own conscience, though it may make that second verification faster. The theological weight of the four conditions tracks the person who says “I believe” and “I testify” to the gathered church, whatever collaborative process produced the words they use to say it.

4. Six Cases

Case One: an AI system generates a sermon and a pastor reads it sight unseen. The words may be doctrinally correct, but the pastor has not verified or appropriated them. The event is best described as human vocalization of generated content, not responsible derivative witness.

Case Two: an AI system generates a full draft; the pastor verifies Scripture use, checks theology, rewrites examples, removes unsupported claims, prays and reflects, and then preaches only what the pastor personally affirms. Here the four conditions can be satisfied. AI has assisted production without becoming the witness.

Case Three: the pastor develops the exegesis and argument while AI improves grammar and organization. This case is even less troubling because core theological judgments originate in the human process and AI functions closer to editorial assistance.

Case Four: a chatbot directly addresses worshipers in the first person, “I testify,” “I believe,” or “I know Christ.” The linguistic form imitates witness while the relevant agency is missing. Unless a human institution explicitly adopts those statements as its own, such language risks a category mistake and possibly deception. Chao’s (2026) FMG-Bench, built to evaluate how large language models handle theological triage and pastoral guidance, is instructive here precisely because it treats the AI as a source of guidance to be evaluated rather than as a witness in its own right; a benchmark that scores a model’s pastoral responses for creedal accuracy, humility, and appropriate referral to a human is doing something closer to quality control on a tool than to assessing testimony. Case Four fails exactly where that framing is abandoned: not when a system gives pastoral guidance well, but when its guidance is delivered in a first-person confessional register that invites hearers to treat the system itself as the believing party.

Case Five: a congregation’s public prayer is drafted with AI assistance and read aloud by a worship leader in the first-person plural, “we confess,” “we thank you.” Prayer differs from preaching in an important respect: its addressee is God, not the congregation, and the congregation’s role is closer to joining an utterance than receiving one. This does not remove the four conditions; it relocates what appropriation and answerability mean. A worship leader who reads an AI-drafted confession without having examined whether its content is something the leader, and the gathered congregation on the leader’s behalf, can actually and sincerely say to God has not appropriated the prayer merely by voicing it well. Because the congregation is invited to join the words as their own, poor verification or appropriation here risks placing false confession or false thanksgiving in the mouths of worshipers who trusted the leader to have already done that work. The stakes of Case Five are not lower than Case One’s; they are different, since the deficient act is directed at God through the assembly rather than at the assembly directly.

Case Six: a pastor uses an AI system exclusively to generate counterarguments and objections to a draft sermon’s central claim, in order to test whether the sermon’s argument survives serious challenge, and then revises the sermon in response without incorporating any AI-generated wording into the final text. No AI language appears in the delivered sermon at all. This case is included because Section 5 names counterargument generation as a legitimate task and no case had tested it directly. Case Six is the least theologically troubling of the six, since the AI’s output never becomes part of the proclaimed content; it functions purely as a sparring partner in the human process of preparation, structurally similar to a colleague who plays devil’s advocate before a sermon is finalized.

5. Objections and Replies

5.1. If it works, why does agency matter?

Reyes’s and Mannerfelt and Roitto’s findings both suggest that audiences can find AI-assisted sermons effective, clear, and even spiritually encouraging. A skeptic might ask why a theological purist’s concern about “true” witnessing agency should carry any weight against congregational or pastoral preference for what evidently works. The reply is that effectiveness and authenticity are answers to different questions. A sermon can be rhetorically effective while raising a live question about whose discernment and responsibility it expresses; the two questions do not compete for the same evidence. Reyes’s own study treats pastoral ownership as a distinct variable from perceived effectiveness precisely because the two can diverge (Reyes 2026). A theology of preaching that collapsed authenticity into effectiveness would have no resources left to distinguish a moving sermon from a moving speech, and Christian proclamation has never accepted that collapse for other reasons, such as plagiarized sermons that are rhetorically effective but ethically compromised.

5.2. Can the framework be gamed?

The four conditions rely substantially on a preacher’s own report of an internal process: verification, assent, appropriation, and answerability are not directly observable by a congregation or an oversight body. A pastor could claim to have done all four while in fact reading an AI draft with minimal engagement, technically satisfying a weak reading of the framework while doing something close to Case One in substance. This is a real limitation and not fully answerable from within the framework alone. Two partial responses are available. First, the framework does not need to be independently verifiable by a congregation to do real theological work; many first-order moral categories, sincerity, repentance, genuine love, are similarly resistant to third-party verification and are not thereby theologically empty. Second, the framework is more useful as a formation tool that shapes how a preacher approaches AI-assisted preparation than as an enforcement tool that a denomination applies after the fact. A preacher who takes the four conditions seriously as a discipline of preparation, asking them of their own process before preaching rather than only being asked them afterward, is less likely to game the framework than one asked to certify compliance retroactively. This response accepts the objection’s force rather than dissolving it: performative satisfaction of the conditions remains possible, and the framework’s value depends on preachers who want to meet the conditions rather than merely appear to.

5.3. Does disclosure resolve the ethical problem?

One might think that full disclosure of AI assistance solves whatever ethical problem attends its use, since hearers would then know exactly what they are receiving. Disclosure and Derivative Witness are related but distinct issues. A pastor could fully disclose that a sermon was AI-drafted and still fail the four conditions, if the disclosed content was never actually verified, assented to, appropriated, or answered for; disclosure would then accurately describe an instance of Case One rather than excuse it. Conversely, a pastor could satisfy all four conditions without disclosing the AI’s role in early drafting, in the same way many pastors do not disclose which commentaries or which colleague’s suggestion shaped a particular illustration. The article’s own position, stated in Section 6 below, is that disclosure requirements should track the risk of hearers misattributing first-person conviction to a source that has none, which is a sharper and narrower standard than a blanket disclosure rule for any AI involvement whatsoever.

5.4. Is the trilemma of three sermon-preparation cases too tidy?

A reader might notice that Cases One through Three form a clean progression from AI origination to human origination, and ask whether real preparation is ever this orderly. It is not, and the article does not claim it is. Sermon preparation typically iterates: a pastor drafts, consults an AI tool, revises, consults again, prays, revises further. The six cases are idealized points along a continuum rather than a taxonomy that exhausts real practice. Their function is to make the four conditions tractable by fixing enough detail to ask, in each case, which conditions are met and which are not; a reader should expect most actual sermons to sit at some unstated point between the cases given, and the diagnostic use of the framework consists precisely in locating a given sermon’s preparation history relative to these fixed points rather than sorting it into one of six bins.

6. Practical-Theological Consequences

Reyes’s experiment is important because pastoral ownership emerged as a limit even where audiences found AI-assisted sermons effective (Reyes 2026). Effectiveness therefore cannot settle authenticity. A sermon may be clear, moving, and doctrinally acceptable while still raising questions about whose discernment and responsibility it expresses.

The model also avoids an unnecessary prohibition. Theological integrity does not require pretending that every phrase in a sermon was produced without tools. What it requires is that the preacher remain the responsible agent of interpretation and proclamation. Disclosure practices may vary by context, but concealment becomes ethically serious when it causes hearers to attribute to the preacher labor, testimony, or first-person experience that the preacher did not in fact undertake.

AI should therefore be evaluated by task. Research assistance, language editing, structural suggestion, and counterargument generation, as in Case Six, can be separated from confessional agency. The more a task bears directly on first-person witness, pastoral judgment, or spiritual counsel, the stronger the need for human appropriation and answerability.

6.1. A disclosure standard

Section 5.3 distinguishes disclosure from Derivative Witness without proposing a positive disclosure standard. One is offered here, stated narrowly enough to avoid a blanket rule that would treat spell-checking and confessional ghostwriting as equally disclosure-worthy. Disclosure is owed when, and only when, a hearer’s reasonable interpretation of the delivered content would attribute to the human speaker a first-person state, such as a personal experience, a private conviction newly reached, or an eyewitness claim, that the speaker did not in fact hold prior to encountering the AI-generated material. An illustration drawn from a pastor’s own life, even if an AI tool suggested how to phrase it, does not trigger this standard, because the underlying experience is genuinely the pastor’s own. A first-person testimony of struggle or conversion that an AI system generated wholesale and that the pastor did not personally live does trigger it, regardless of how thoroughly the pastor otherwise verified, assented to, and appropriated the language, because the content itself falsely claims a lived origin. This standard is deliberately narrower than “disclose all AI use,” which this article does not defend, and deliberately broader than “disclose nothing,” which the concealment risk in Section 6 rules out.

6.2. Formation and training implications

If verification, assent, appropriation, and answerability are skills as much as dispositions, seminary and continuing-education formation bears some responsibility for cultivating them, not merely assuming pastors will exercise them correctly once ordained. A curriculum that trains students to use AI tools for research and drafting without also training the specific discipline of checking, owning, and defending what results risks producing exactly the failure mode this article warns against, competent tool use without accountable authorship. Practically, this suggests homiletics courses might include supervised exercises in which students receive an AI-generated draft and are assessed not on the draft’s quality but on the rigor of their verification of it, the specificity of their reasons for assent or dissent to particular claims, the extent of their appropriation as measured by how much the delivered sermon diverges from the draft in voice and application, and their capacity to defend contested claims in a mock question-and-answer setting afterward. Such an exercise would train the four conditions directly rather than leaving them to be acquired incidentally.

Practice-based ministry projects are already attempting something like this at the level of a single congregation. Poe’s (2026) doctoral project develops a Christ-centered approach to AI-assisted sermon preparation for one specific church, rather than a general theological account of the kind offered here. Local, congregation-specific guidance of this sort is a natural complement to the present framework rather than a competitor to it: a denomination or seminary might reasonably adopt something like the four conditions as its general theological standard while leaving the concrete procedures, which tools are approved, what a verification checklist looks like, how associate staff are trained, to exactly the kind of local, practice-based development Poe’s project represents. The present article does not evaluate Poe’s specific procedures, which were designed for one congregation’s context and are not reported here, but it notes the convergence of independent projects on the same underlying concern as one further reason to think the problem this article addresses is not a merely academic one.

7. Limits and Research Agenda

The argument here is conceptual rather than empirical. It does not measure how frequently pastors currently meet or fail the four conditions in actual practice, nor does it establish that congregations can detect the difference between Derivative Witness and its failure modes when they encounter them. Reyes (2026) and Mannerfelt and Roitto (2025) provide the closest available empirical grounding, but neither study was designed to test the specific four-condition framework proposed here, and both concern sermon preparation rather than prayer or pastoral counsel.

Several questions follow directly. First, can verification, assent, appropriation, and answerability be operationalized as observable indicators for empirical study, for instance through structured interviews or think-aloud protocols during sermon preparation, without reducing them to a checklist that misrepresents their character as dispositions rather than discrete acts? Second, do congregations that learn, after the fact, that a valued sermon involved substantial AI drafting revise their theological assessment of that sermon, and does this revision track the four conditions or something else entirely, such as a general discomfort with the technology regardless of how it was used? Third, how does the analysis offered here for individual preachers extend to institutional proclamation, official denominational statements, catechetical materials, or liturgical texts produced with AI assistance and adopted by a body rather than delivered by a single accountable speaker? The present article’s account of collective derivative witness in Section 3.3 is a first approximation and not a developed answer to this third question. Fourth, does the sincerity-condition account of witness developed in Section 2.1 generalize to non-Christian traditions with their own accounts of testimony and proclamation, or does it depend on theological commitments specific to the Christian doctrine of witness that this article has assumed rather than defended? These questions mark the boundary of what a conceptual article of this kind can establish and the beginning of what further theological and empirical work would need to take up.

8. Conclusion

AI can produce proclamation-content. That does not make AI a Christian witness. The theological act of witness includes an agent who stands behind the speech through belief, intention, and accountability, and speech-act theory’s sincerity condition gives that requirement a precise philosophical shape rather than leaving it as an intuition.

Derivative Witness offers a practical middle category. AI-generated language can become part of a human act of proclamation when a human witness verifies, assents, appropriates, and accepts answerability for it, whether fully or in the partial and mixed ways real preparation actually allows. The decisive question is therefore not “Was AI used?” but “Who, in the end, owns this claim as witness?”