AI Systems Encode the Worldviews of Fallen Programmers, Not Neutral Facts

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

The Gist

The model is not a blank window onto truth. People with worldviews built it, people with worldviews filled it with data, and people with worldviews decided what it is allowed to say. Read it like that, not like a referee in the sky.

Conclusion

There is no such thing as a theologically or morally neutral AI system, because there are no neutral programmers, no neutral datasets, and no neutral definitions of success. Whoever creates the content determines the narrative. Decisions about what to surface, what counts as harm, misinformation, or hate speech, and what a “helpful” answer is, are philosophical and theological decisions whether or not the programmers admit it. Christians should therefore examine outputs against Scripture rather than treating the system as an oracle.

Premises

1. Every line of code and every training choice is made by people of whom Romans 3:23 is true. Neutrality is impossible when fallen human beings do the programming, the training, and the prioritizing. Walker’s pastoral test: attempts to push a strictly biblical worldview through these systems often meet refusals or error messages. The right question is not only “what did it say?” but “why was it unable?”
2. Bishop Charlie Hames Jr.’s line, “whoever creates the content determines the narrative,” applies. If an AI system looks objective, that appearance is itself part of the narrative its makers encoded. Ranking, filtering, and “safety” are someone’s moral framework running at scale.
3. Harrison cites Statista-style figures that a large share of AI’s cited or training-adjacent sources are platforms such as Reddit, Wikipedia, YouTube, and Google. Even if the exact percentages are contestable, the steelman does not need them to be precise. It needs only that dominant corpora are not a random sample of biblical wisdom. They disproportionately reflect secular and progressive discourse, including material hostile to Christian moral claims.
4. AI is therefore philosophical, ideological, and even theological. It is not a vending machine that indiscriminately spits facts. Thomas Brooks’s image applies: the bait can hide the hook. Apparent usefulness can smuggle a moral formation.

Assumptions

• “Bias” here is not a claim that every answer is a lie. It is a claim that selection, ranking, and refusal policies are value-laden.
• Christians may still use the tools (Walker says he uses them a great deal) provided they check outputs against original sources and Scripture.
• 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 53:34–2:18:08 of the episode; it is not a verbatim transcript.

Analysis

Overall strength: Moderate. Argument type: Deductive.

Premise Strength

• P1: Fallen programmers / pastoral test of refusals (Weak) — The theological claim that programmers are morally fallen is a matter of accepted doctrine within the argument's framework and is not itself contested. But the empirical payload — that this fallenness manifests as detectable, targeted refusal of biblical content — rests on a single, undocumented anecdotal pattern with no controls, sample size, or comparison to how the same systems treat other strong ideological content.
• P2: 'Whoever creates the content determines the narrative' (Hames) (Weak) — This functions largely as a restatement of the conclusion in aphoristic form rather than as independent support. It is conceptually true in a broad sense (curatorial power exists) but treats AI companies as unified, coherent ideological agents, glossing over internal pluralism, legal/liability drivers, and competitive dynamics that also shape outputs.
• P3: Statista-style figures on Reddit/Wikipedia/YouTube dominance in training-adjacent sources (Moderate) — This premise identifies a real and checkable phenomenon (corpus composition skews toward a narrow set of platforms) and is appropriately hedged regarding precision. However, the further inferential leap—from corpus composition to demonstrated hostility toward Christian moral claims—is not established, since post-training alignment processes are widely regarded as more proximate determinants of model behavior than raw pretraining statistics.
• P4: AI is philosophical/theological, not a neutral vending machine (Weak) — This operates as rhetorical restatement and framing (the bait-and-hook metaphor) rather than independent argumentative support; it draws its persuasive force from P1–P3 rather than adding new evidence.

Potential Fallacies

• Suppressed premise / enthymeme (Inference from P1 to the Conclusion) — The argument treats the leap from 'programmers are morally fallen' to 'therefore AI outputs are non-neutral in a targeted, meaningful sense' as self-evident, but this requires an unstated bridging premise (that fallenness reliably transmits into artifact-level bias of a particular ideological direction) that is never defended and does substantial hidden work.
• Hasty generalization / anecdotal overweighting (P1 (Walker's pastoral test)) — The 'pastoral test' generalizes from an unspecified, undocumented number of refusal incidents (no prompts, sample size, model version, or comparison condition given) to a systemic claim about AI's theological hostility. Without testing whether equally strong secular, Islamic, or other dogmatic prompts trigger similar refusals, there is no way to distinguish targeted bias from generic guardrails against absolutist or dogmatic phrasing of any kind.
• Unfalsifiable framing (P1 and P2) — The argument's structure allows any AI behavior to confirm the thesis: apparent neutrality is reinterpreted as 'part of the encoded narrative' (P2), while refusals are read as evidence of suppression (P1). Because no possible output could disconfirm the claim, it functions more as an unfalsifiable interpretive lens than a testable empirical hypothesis.
• Equivocation between weak and strong claims (motte-and-bailey) (Transition from A1/P1/P2 (weak claim) to P3/P4 and the Conclusion (strong claim)) — The argument moves fluidly between a nearly trivial, well-supported claim (all information systems involve value-laden selection) and a much stronger, contested claim (dominant AI systems are disproportionately hostile to Christian moral claims specifically). The premises establish the weak claim; the conclusion and rhetorical framing assert the strong one, without additional evidence bridging the gap.
• Weak inductive leap / missing causal mechanism (P3 to P4) — The argument infers output-level ideological bias directly from pretraining-corpus composition (Reddit, Wikipedia, YouTube), but modern AI behavior — especially refusals and framing — is shaped more directly by post-training alignment processes (RLHF, safety classifiers, system prompts) than by raw source statistics. This intermediate causal layer is never addressed, weakening the inference from P3 to P4.
• Proves too much (self-undermining asymmetry) (General form of P1/P4, applied to the Conclusion and A2) — If moral fallenness in creators guarantees non-neutral, ideologically compromised output, this applies equally to Bible translation committees, sermons, commentaries, and the podcast's own theological claims — including the very process of 'checking against Scripture' that the argument recommends as a corrective. The argument offers no principled reason why its preferred epistemic authority is exempt from the critique it levels at AI.
• Appeal to authority without independent corroboration (P2 and P4) — Quotations from Bishop Hames and Thomas Brooks are rhetorically evocative but function as restatements of the thesis in memorable language rather than as independent evidence that AI systems behave in the specific way alleged.

Counterarguments

• P1 (High impact) — AI refusal behavior is well documented to occur across many types of dogmatic, absolutist, or strongly normative prompts regardless of ideological content (political, religious, or otherwise), reflecting generic safety heuristics rather than targeted theological suppression. Without a controlled comparison across worldviews, the 'pastoral test' cannot distinguish these hypotheses.
• P3 to P4 inference (High impact) — Model behavior is shaped predominantly by post-training alignment (RLHF, constitutional AI, system prompts) rather than raw pretraining corpus composition; several AI labs have published documents acknowledging value-laden design choices while also implementing measures intended to avoid taking sides on contested religious or political claims. This undercuts the direct causal chain from 'secular-leaning training data' to 'output hostile to Christian claims.'
• Conclusion / A2 (High impact) — The same 'fallen humans cannot produce neutral content' logic applies equally to Bible translation committees, sermons, denominational teaching, and the podcast's own argument, since these are all produced by fallen humans with value commitments. This creates an unaddressed asymmetry: the argument exempts its own preferred epistemic authority from a critique it treats as universally binding on AI.
• Overall thesis novelty (Medium impact) — The claim that no AI system is fully neutral is now close to consensus among AI developers and ethicists themselves (many have explicitly disclaimed pure neutrality in public model documentation), which recontextualizes the argument's central claim as less controversial and less distinctively theological than its framing suggests.
• P3 (Medium impact) — The Reddit/Wikipedia/YouTube skew, if accurate, more plausibly reflects the demographics of English-language internet contributors generally (educated, online-active populations) than a deliberate or even emergent anti-Christian hostility; 'secular-leaning' and 'actively hostile to Christian claims' are conflated without justification.

Suggested Improvements

• Empirical rigor for P1 — Replace the anecdotal pastoral test with a systematic, controlled comparison: submit matched sets of strongly-worded prompts across multiple worldviews (biblical, secular-humanist, Islamic, Marxist) to the same systems and compare refusal rates. This would allow the argument to distinguish targeted theological bias from generic guardrails against dogmatic or absolutist framing, which is currently the argument's single greatest evidentiary vulnerability.
• Causal specificity for P3-P4 — Address the distinction between pretraining corpus composition and post-training alignment (RLHF, system prompts, safety classifiers), and explain why corpus statistics should be expected to predict output-level bias despite this intervening layer. Without this, the argument's central empirical bridge (source composition → output hostility) remains an unsupported inferential leap that domain experts would immediately challenge.
• Self-consistency / reflexivity — Explicitly address why the fallenness-implies-non-neutrality principle does not equally undermine Scripture-reading, translation, and pastoral interpretation, or concede that all mediated knowledge (including the recommended corrective) requires similar scrutiny. This closes the most damaging reductio available to critics and strengthens the argument's internal coherence without requiring abandonment of its core theological claim.
• Scope clarity — Separate explicitly the modest, well-supported claim (all AI systems involve value-laden choices) from the stronger, contested claim (current mainstream AI is specifically biased against Christian moral claims), and mark the confidence level appropriate to each. This would prevent the rhetorical slide from a claim nearly everyone can grant to a claim requiring far more evidence, improving both persuasiveness to skeptical audiences and intellectual honesty.

Scenario Tests

• Controlled testing shows AI systems refuse or hedge on strongly-worded prompts from many worldviews (atheist, Islamic, secular-humanist, biblical) at similar rates. (Challenges) — This would defuse P1's rhetorical force ('why was it unable') by showing the refusal pattern reflects content-agnostic guardrails against dogmatic assertion rather than theological targeting, reducing the argument to the much weaker and less actionable claim that 'nothing is perfectly neutral.'
• The same fallenness-based skepticism is applied reflexively to Bible translation committees, denominational curricula, and the podcast's own theological claims. (Challenges) — This generates a reductio: if fallen human mediation guarantees non-neutral, potentially untrustworthy output, then the recommended corrective (checking AI against Scripture as interpreted by fallen humans) is equally suspect, undermining the argument's practical conclusion.
• AI systems are shown to readily generate accurate biblical content, scripture citations, and robust Christian apologetics without resistance in the vast majority of interactions. (Challenges) — This would complicate the narrative of systemic hostility to Christian moral claims, suggesting that any observed friction is topic-specific or prompt-specific rather than indicative of a pervasive anti-Christian ideological architecture.
• Published training-data audits or model specification documents confirm secular/progressive-leaning source skew and show measurable correlation with output framing on contested moral topics. (Supports) — This would strengthen P3-P4's causal claim considerably, providing the kind of systematic evidence currently missing and partially validating the argument's stronger claim, though it would still need to rule out generic (non-Christian-specific) explanations.

Coherence & Relevance

The argument is internally coherent as a piece of theological rhetoric and its foundational claim — that AI systems embed value-laden choices and cannot achieve pure neutrality — is well-supported and increasingly uncontroversial. However, the argument's practical thrust depends on a stronger, directional claim (systematic bias against Christian moral claims) that is not adequately supported by the anecdotal and hedged-statistical evidence offered. The premises also do not address the most damaging challenge to the argument's internal consistency: that its own fallenness-based skepticism, if applied reflexively, would equally undermine the Scripture-checking corrective it recommends. The argument would be substantially strengthened by controlled comparative evidence, engagement with the alignment/RLHF layer of AI development, and explicit acknowledgment of its own reflexivity problem.

• P1: Fallen programmers / pastoral test (Moderate) — The theological premise (Romans 3:23) is relevant to the philosophical claim that neutrality is impossible in principle, but the empirical anecdote offered as corroboration does not adequately connect to the specific, directional claim of anti-Christian bias without a comparison condition.
• P2: 'Whoever creates the content determines the narrative' (Weak) — This functions as a restatement of the conclusion rather than independent support; it does not add new evidence connecting programmer identity to specific output patterns.
• P3: Corpus composition statistics (Moderate) — Relevant to establishing that training data is not demographically representative of biblical wisdom, but the premise does not bridge to the stronger claim of output-level hostility without addressing the alignment/RLHF layer that more directly governs model behavior.
• P4: AI as philosophical/theological, not neutral vending machine (Weak) — Largely rhetorical restatement of the conclusion using metaphor (bait and hook); it doesn't independently advance the case beyond what P1-P3 already assert.

View this argument on LogicFirst.ai