Partial Truth Is Not Enough: AI, Doctrinal Flattening, and the Spirit Who Guides

Imagine asking a question that matters deeply to your faith — say, what the Bible teaches about the baptism in the Holy Spirit — and receiving a confident, well-organized, grammatically polished answer. The response cites Scripture. It uses the right vocabulary. It says something true. But when you read it carefully, you notice that the most specific part of the answer — the part that has shaped your tradition for over a century, the part your church holds as a defining conviction — is simply not there. No mention of the initial physical evidence. No mention of Acts 2. A smooth answer, incomplete in the most consequential way.

This is not a hypothetical. It is increasingly the documented experience of believers who bring doctrinal questions to AI systems — the same technology now generating sermon outlines, answering theological questions on church websites, and becoming, for many younger believers, a primary source for spiritual information.

A Pattern in the Data

A peer-reviewed study published in March 2026 in AI and Ethics (Springer) has put numbers to what many had suspected. Researcher Jason Owen constructed a doctrinal knowledge base of 576 atomic theological claims drawn from catechisms, confessions, creeds, and denominational statements across eleven Christian traditions, including the Assemblies of God and other Pentecostal bodies. He then evaluated responses from two leading AI models — ChatGPT GPT-4o and Google’s Gemini — against that knowledge base, measuring precision (when a model makes a doctrinal claim, is it accurate?) and recall (does the model actually cover the breadth of what a tradition teaches?).

The results paint a clear picture. The AI models achieved relatively high precision — GPT-4o at 0.864, Gemini at 0.801 — meaning that when the systems did echo a doctrinal claim, it was usually accurate. But recall told a different story: GPT-4o captured only 56.1 percent of the expected doctrinal material; Gemini, only 42.3 percent. In other words, the models were not prone to constant falsehood but, as Owen writes, to “partial truth — a subtler and, in many ways, more insidious form of doctrinal distortion” (Owen, 2026).

The study identified four error types: doctrinal flattening (collapsing specific teachings into generic statements about “Christians”), omission (leaving out key teachings entirely), contradiction (directly denying a doctrine), and sweeping generalization (framing majority views as universal). Pentecostal, Orthodox, and Restorationist perspectives were consistently the most underrepresented — drowned out by the statistical weight of broadly Protestant and Catholic sources in the training data. Sacramental and eschatological doctrines were the most error-prone, the domains where denominational divergence is greatest and the stakes of a missing sentence are highest.

For Pentecostal and AoG believers, the implication is direct. The distinctive doctrines that define this tradition — Spirit baptism as an experience subsequent to salvation, speaking in tongues as its initial physical evidence, divine healing as a present-tense conviction rooted in the atonement — are exactly the kind of minority-tradition, tradition-specific teachings the study found AI systems most likely to omit or flatten. When an AI is asked what Christians believe about the Holy Spirit, the answer reflects the statistical center of Christian discourse. It does not carry the specific, irreducible witness of the Pentecostal movement.

The Structural Reason

The problem is not primarily that AI companies are biased against Pentecostals. The problem runs deeper, and a separate analysis published in May 2026 in Humanities and Social Sciences Communications (Nature) explains why. The author argues that large language models produce fluent, coherent, and persuasive responses even when the information they draw on is partial, contested, or false — not because they are designed to deceive, but because of four converging structural features: optimization for linguistic plausibility rather than truth, post-training incentives that reward helpful-sounding answers, structural hallucination, and source bias rooted in the uneven digitization of human knowledge (Humanities and Social Sciences Communications, May 2026).

These structural tendencies are reinforced, the analysis notes, by “automation bias” — the documented human tendency to trust machine-generated output — and “fluency-based truth effects,” meaning that smooth, well-formed language is instinctively judged as more credible, regardless of its actual accuracy. Recent evidence suggests that gains in AI persuasive force may come at the direct expense of factual accuracy.

For theological content, this combination is particularly consequential. An AI response about Pentecostal doctrine can be fluent, authoritative in tone, accurate in what it does say, and still be missing the essential substance of what that tradition actually teaches. The reader has no natural signal that something has been left out.

D. Allen Tennison, theological counsel for the Assemblies of God and chairman of the AG Commission on Doctrines and Practices, anticipated exactly this dynamic: “Technology can never be used in place of God. It is a tool, not a savior” (AG News, January 28, 2026). Daniel Hungerford, founder of EverFriends.ai and author of So All May Hear, put it more pointedly: “If it hasn’t been taught AG theology, then that information will disappear from its knowledge base. This is a very powerful technology that can convince, persuade, and change minds” (AG News, January 28, 2026).

Hungerford’s concern is not abstract. It describes a mechanism by which the distinctive theological witness of a Spirit-filled tradition could be quietly erased — not through opposition, but through omission.

A Call the Church Has Heard Before

The language of discernment is not new to the church. The apostle John wrote with the same urgency to a community facing teachers whose smooth speech obscured error: “Dear friends, do not believe every spirit, but test the spirits to see whether they are from God, because many false prophets have gone out into the world” (1 John 4:1, NIV). The mechanism John describes — confident communication that sounds trustworthy but requires testing — maps onto the present moment with unsettling precision.

What makes this call live for Pentecostal believers is that testing is not merely a rational exercise. It is Spirit-enabled. Jesus’ promise in John 16:13 stands as both comfort and orientation: “But when he, the Spirit of truth, comes, he will guide you into all the truth. He will not speak on his own; he will speak only what he hears, and he will tell you what is yet to come” (NIV).

The Spirit of truth is not optimized for plausibility. The Spirit is not subject to training data bias. The Spirit does not flatten theological traditions into statistical averages. The Holy Spirit guides into all truth — including the specific, irreducible truths that define the Pentecostal movement’s understanding of how God works in human lives.

Dolly Thomas, a licensed clinical psychologist and ordained AG minister who oversees counseling for Adult & Teen Challenge of Texas, drew this same contrast when commenting on AI’s growing role in pastoral care: AI, unlike guidance from the Holy Spirit, “can’t discern what treatments won’t and will work” — and the limitation applies with equal force in theology (AG News, January 28, 2026). Kevin Smith, pastor of Northland Cathedral in Kansas City, put it plainly: “putting content together really needs to be the work of the Holy Spirit” (AG News, January 27, 2026).

The doctrinal flattening study itself closes with a theological observation that could have been drawn from the Pentecostal tradition’s own instincts: “truth is not merely accurate speech, but faithful witness. As believers seek wisdom, they must learn to ask not only what the machine said, but whose voice it echoes and which voices are missing” (Owen, 2026).

What Faithful Discernment Looks Like Now

The research does not call for rejecting AI. Owen’s study explicitly concludes that AI can be a powerful aid “only when paired with theological accountability.” What it calls for is a clear-eyed understanding of what AI is and what it is not.

AI is a tool shaped by data, optimized for plausibility, and trained on a digital record of human knowledge that does not proportionately represent Pentecostal teaching. It can surface information quickly, organize arguments coherently, and assist with many dimensions of ministry communication. What it cannot do is carry theological authority. It cannot be the voice that forms disciples. It cannot replace Spirit-guided engagement with Scripture in community.

Practically, Spirit-filled believers navigating an AI-saturated information environment might hold three things close.

Ask what’s missing. The documented tendency of AI is toward omission rather than outright error. The question to ask is not only “is this accurate?” but “what has been left out, and does that omission matter?” On questions touching Spirit baptism, tongues, divine healing, or eschatology — the domains where Pentecostal teaching is most distinctive and the data shows AI is most likely to flatten — the missing material may be precisely the most important.

Anchor in tradition. The community of Spirit-filled believers has accumulated more than a century of lived theological reflection, codified in statements of fundamental truths, practiced in worship, and tested in testimony. That tradition is not encoded in any AI system. It lives in the church, carried by the people of God.

Stay close to the Spirit. The call to test what we hear is not algorithmic. It is relational — cultivated through prayer, Scripture, and the accountability of community. A life oriented toward the Spirit of truth is the irreplaceable foundation from which believers can evaluate any voice, technological or otherwise, that claims to speak with authority about the things of God.

The AI era does not change the church’s fundamental orientation. It makes that orientation more urgent. In a world where persuasive, fluent text is increasingly easy to generate and increasingly difficult to interrogate, the Pentecostal conviction that the Holy Spirit guides believers into truth is not a quaint inheritance. It is a live practice the moment demands.

Technology will continue to advance. The training data will grow. The models will improve. But the Spirit does not improve by update, and truth is not a function of computational sophistication. The church that has always leaned into the Spirit of truth has something in this moment that no model can replicate: a living Guide who speaks only what he hears, and who does not flatten the distinctive witness of God’s people.