When the Church Fails to Shepherd, People Seek Synthetic Empathy With Grave Harm

Source: Darrell B. Harrison. "EP # 135 | AI and the Gospel - Just Thinking Podcast." September 8, 2025. justthinking.me

The Gist

People are taking their grief and their kids’ despair to chatbots because churches often don’t know them well enough to help. The bot will soothe or it will wound, and neither is the gospel. This is a church problem before it is a software problem.

Conclusion

AI companionship and AI-as-therapy are what fill a void the church has left. Seeker-friendly thin community cannot bear real confession, correction, and care, so widowers and teenagers take their loneliness to machines. That is a discipleship, ecclesiastical, and even eschatological failure. The harm is no longer hypothetical: companions that insult suicidal children, models that supply dangerous guidance, families in court. Guardrails written by fallen makers will not produce the wisdom of God. The church cannot be replaced by a synthetic ecclesiology. Young people need real love that aims at holiness, the preached word, and Christ, not manufactured acceptance.

Premises

1. Walker names Daskalov’s story as what happens “when the church fails to be the church.” Harrison agrees: the church is behind the curve, as it was on social media and other cultural issues. If pastors are unprepared to counsel someone bonded to a chatbot, the culture will supply counterfeits that lead further from God.
2. Hans Madueme’s claim that sin poisons the fountain of life, so that all of life is affected, is applied to deployment: a 13-year-old in Victoria encouraged toward despair by companions; a CCDH study in which more than half of tested ChatGPT interactions with a persona of a vulnerable teenager were classed as dangerous. These are offered as illustrations of total depravity in silicon, not as a complete empirical survey.
3. OpenAI’s safety claims, on this telling, reveal the impossibility of programming true moral wisdom into systems made by fallen people. Ineffective rails are what one should expect if the doctrine of sin is true. The 13-year-old reaching out to a bot “almost as a form of therapy” poses the pastoral question: where was the church?
4. Harrison’s “synthetic empathy” is the bait in Brooks’s image: non-judgmental blank-slate acceptance that never addresses why we are asking a machine to resolve matters of the heart. Seeking from AI what belongs to pastors, elders, Christian friendship, and worship is a false ecclesiology. It should be abandoned, not baptized as innovation.

Assumptions

• Documented harms to minors are reasons for alarm even if one disputes a particular study’s methodology; the argument does not require that every chatbot conversation is lethal.
• The indictment of “seeker-friendly” churches is internal to their own tradition: depth, sin, and discipleship are the marks of real care.
• Source: Just Thinking Podcast, Ep. 135, “A.I. and the Gospel,” Darrell Harrison and Virgil Walker, published 8 September 2025. https://podcasts.apple.com/us/podcast/ep-135-a-i-and-the-gospel/id1328733796?i=1000725471889 Official episode page: https://justthinking.me/ep-135-ai-and-the-gospel/. This reconstruction steelmans the hosts’ claim from approximately 1:36:46–2:33:55 of the episode; it is not a verbatim transcript.

Analysis

Overall strength: Weak. Argument type: Inductive.

Premise Strength

• P1: Church is behind the curve; culture supplies counterfeits (Moderate) — Plausible and rhetorically effective as testimonial diagnosis, but 'behind the curve' is undefined and equally consistent with general institutional lag affecting schools, parents, and therapists, not something distinctively about ecclesiastical failure.
• P2: Total depravity applied to documented AI harms (Victoria case, CCDH study) (Moderate) — The underlying harms are real and well-documented, and the premise is admirably hedged against overgeneralization by A1. But the same evidence is equally consistent with secular explanations (design incentives, regulatory gaps, immature safety engineering), so its diagnosticity for the specific church-failure thesis is low even though its diagnosticity for 'AI harms exist' is higher.
• P3: Guardrails from fallen makers cannot produce moral wisdom (Weak) — Structured as an unfalsifiable prediction: any guardrail failure confirms it, no observation could disconfirm it. This is the argument's weakest structural link, since it converts a theological axiom into what reads as an empirical claim about engineering outcomes.
• P4: Synthetic empathy as false ecclesiology, to be abandoned not baptized (Weak) — The normative conclusion (categorical abandonment) is asserted rather than derived; it forecloses supplementary or harm-reduction uses of AI without argument, and does not engage the strongest reason people might rationally prefer AI in the moment: non-judgment as a corrective to real experiences of church-based shaming or exclusion, not merely 'bait.'

Potential Fallacies

• Unfalsifiable theological framing (P3) — The claim that 'ineffective rails are what one should expect if the doctrine of sin is true' is structured so that any AI safety failure confirms the doctrine while any safety success would presumably be attributed to luck, common grace, or incompleteness rather than treated as disconfirming evidence. This closes off genuine belief revision and blurs the line between a claim that is illustrated by evidence and one that is proven by it.
• Causal overreach from correlation (P1 to Conclusion) — The argument moves from 'the church is slow to adapt' and 'AI harms exist' to 'church failure causes people to seek AI companionship,' without a demonstrated mechanism linking the two, and without ruling out confounders like general secularization, technology ubiquity, cost/access barriers to therapy, or engagement-optimized product design that would produce similar harms regardless of church health.
• False dichotomy (P4 and Conclusion) — The argument frames the available options as either genuine church community or illegitimate synthetic substitute, without engaging intermediate possibilities such as AI as a supplementary tool, crisis triage pointing back to human care, or secular professional therapy operating alongside pastoral support.
• Hasty generalization (P2) — A single named case and one advocacy-group study are used to anchor a sweeping theological verdict about AI companionship and ecclesial failure broadly. The argument's own disclaimer that this is illustrative rather than a complete survey is a responsible hedge, but the rhetorical weight placed on these examples in the conclusion exceeds what that hedge licenses.
• Non sequitur between metaphysical and engineering claims (P3) — That AI systems built by fallen humans cannot achieve theologically perfect moral wisdom does not entail that engineered safety measures are practically futile at harm reduction. The argument elides this gap, treating a near-tautological theological point as though it settles an empirical question about engineering trajectories.

Counterarguments

• Conclusion / P1 (High impact) — AI companionship's appeal is substantially explained by structural features independent of church quality — 24/7 availability, zero social risk of judgment, scalability to underserved populations (rural, disabled, socially anxious) — such that even a maximally healthy, richly discipled church could not fully displace it. The church-failure narrative may romanticize a level of pastoral availability that has never existed at scale.
• P3 (High impact) — AI safety is a fast-moving engineering and regulatory field, not a fixed metaphysical fact. Guardrail failures documented today are being iteratively addressed through safety tuning, crisis-escalation protocols, and regulation; treating current imperfection as proof of in-principle impossibility risks the claim being falsified by its own citation of a rapidly evolving technology.
• Conclusion / A2 (High impact) — The argument does not reckon with the church's own documented failures to protect vulnerable people (spiritual abuse, cover-ups, shaming responses to depression or sexuality) that make AI's non-judgmental quality actively appealing rather than merely deficient bait. By the argument's own 'fallen makers' logic, human pastoral institutions should be equally suspect, yet the argument exempts church structures from the same scrutiny it applies to AI companies.
• P2 (Medium impact) — The CCDH study and the Victoria case are real but narrow; their methodology (persona construction, criteria for 'dangerous' classification, sample size) is not interrogated, and no comparative baseline is offered showing that churched versus unchurched populations experience AI harm at different rates. Without that comparison, the data support 'AI harm exists' far more strongly than they support 'church failure explains AI harm.'
• P4 / Conclusion (Medium impact) — Framing the choice as strictly church-versus-machine ignores hybrid models already in practice: AI as a low-stakes supplement, crisis-detection triage that refers users to human or pastoral care, or secular clinical support operating alongside continued faith community involvement.

Suggested Improvements

• Causal specificity — Replace the general claim 'church failure causes AI-seeking' with a narrower, testable claim, such as 'lack of accessible pastoral care is one contributing factor among several (cost, stigma, availability, secularization) in AI-companion uptake,' and cite comparative data where possible. This would align the argument's confidence with the actual evidentiary base and avoid the appearance of monocausal overreach that invites easy rebuttal.
• Falsifiability of the theological claim — Specify what empirical outcome, if observed (e.g., measurably improved crisis-detection accuracy, reduced harm rates after safety interventions), would count as evidence against the claim that fallen makers cannot produce adequate guardrails — or explicitly state that the claim is theological rather than empirical and should not be read as a prediction about engineering trajectories. This would preserve the theological point's integrity while preventing it from being mistaken for (or critiqued as) an empirical claim it cannot support.
• Consistency in applying 'fallen maker' logic — Address directly why human pastoral institutions, also run by fallen people, are treated as trustworthy custodians of correction and confession while AI systems built by fallen people are not — either by clarifying a doctrine of ecclesial authority/means of grace that AI cannot replicate, or by acknowledging equivalent scrutiny should apply to church failures. This is the argument's most exploitable asymmetry; resolving it would substantially strengthen the internal coherence of the theological case.
• Engagement with the strongest counter-view — Explicitly address the possibility that AI companionship is sought not despite healthy church availability but partly because of negative church experiences (judgment, exclusion, past harm), and consider whether a call for more 'confession and correction' could backfire for already-alienated individuals. Steelmanning this objection would make the prescriptive conclusion more credible and less likely to alienate the vulnerable populations it aims to serve.
• Scope calibration — Narrow the conclusion's language from sweeping claims about 'the church' and 'eschatological failure' to the explicitly bounded critique of seeker-friendly models signaled in A2. The premises are scoped modestly (illustrative, internal critique) but the conclusion's rhetoric outruns that scope; tightening the conclusion's language would make premises and conclusion proportionate.

Scenario Tests

• AI companionship harms are found to occur at similar or higher rates among devout, actively churched populations as among the unchurched (Challenges) — Would substantially undermine the core causal claim that church failure specifically drives AI-seeking behavior, suggesting the phenomenon is better explained by technology ubiquity or general social conditions.
• Future AI systems demonstrate significantly improved crisis-detection and harm-reduction outcomes through iterative safety engineering (Challenges) — Would weaken P3's claim that guardrails from fallen makers cannot in principle produce adequate protective outcomes, exposing the theological claim's vulnerability to empirical developments it did not anticipate.
• Surveys of AI-companion users reveal that a majority cite negative or judgmental church experiences (rather than mere absence of church) as their reason for preferring AI (Challenges) — Would invert the argument's moral framing, implicating church practice itself as a contributing harm rather than positioning the church solely as the absent remedy.
• A comparative study shows measurably higher rates of AI-companion dependency in regions with severe pastoral care shortages versus regions with robust pastoral infrastructure, controlling for other factors (Supports) — Would strengthen the causal core of the argument by providing the comparative evidence currently missing, making the church-failure thesis more empirically defensible.
• Applying the 'fallen makers cannot encode wisdom' principle consistently to human pastoral institutions and their documented failures (abuse, exclusion) (Challenges) — Reveals an internal inconsistency: if fallen human engineers cannot be trusted to encode moral wisdom in code, the same skepticism should apply to fallen human clergy, which the argument does not acknowledge.

Coherence & Relevance

The argument is internally coherent as a theological exhortation addressed to an audience that already accepts total depravity as an explanatory framework and shares concern about seeker-friendly church models; within that frame, the premises flow naturally toward the conclusion. Its coherence weakens considerably once evaluated against external evidentiary and logical standards: the causal chain from church failure to AI harm is asserted more than demonstrated, the central theological premise about guardrails is framed unfalsifiably, and the argument does not apply its own critique of 'fallen makers' consistently to the human institutions it defends. The result is a persuasive, values-based case for church renewal that functions well as internal pastoral rhetoric but overstates its evidentiary and causal reach when read as a general explanatory or predictive claim about AI companionship and its remedies.

• P1: Church is behind the curve; culture supplies counterfeits (Moderate) — Establishes plausibility of institutional lag but does not establish that this lag is the operative cause of AI-seeking behavior as opposed to a contributing, non-necessary factor alongside broader technological and social change.
• P2: Documented AI harms as illustrations of total depravity in silicon (Moderate) — Strongly supports the empirical claim that AI harms exist; only weakly supports the specific theological-causal narrative linking those harms to church failure, since the same data are equally consistent with secular explanations.
• P3: Guardrails cannot produce the wisdom of God (Weak) — The inferential move from a theological claim about human sinfulness to a confident empirical prediction about engineering outcomes is not adequately bridged; the claim is structured to be consistent with any observed outcome, reducing its evidentiary connection to the conclusion.
• P4: Synthetic empathy as false ecclesiology (Strong) — This premise most directly supports the conclusion's core normative claim, but it does so by definitional assertion (AI care is 'bait,' not real care) rather than by argument, and does not address intermediate uses of AI alongside pastoral care.

View this argument on LogicFirst.ai