AI Misquotes the Bible: A Theology of Truth-Telling

Even the best AI models misquote Scripture at least 15 percent of the time — and some get it wrong as often as 60 percent — according to YouVersion CEO Bobby Gruenewald, whose Bible app serves over a billion installs (1). This is not a reason for Christians to abandon AI; it is a reason to use it under a discipline the church has carried for three thousand years: the discipline of truth-telling. The machine supplies fluency. The believer must supply true claims — grounded sources, verification before quotation, and a human being who signs his name to every word.

Key Takeaways

• YouVersion CEO Bobby Gruenewald reports that the best AI models misquote Scripture at least 15% of the time, with some models erring up to 60% of the time — which is why YouVersion has refused to ship a theological chatbot until accuracy is provable (1)(2).
• AI systems have been documented fabricating Bible verses outright — inventing plausible-sounding text and attaching real chapter-and-verse references to it (3).
• Hallucination is not a glitch but a design property: large language models are optimized to produce plausible text, not verified truth, which makes them fluent and confident even when wrong (4)(5).
• Grounding AI in approved knowledge works: retrieval-augmented generation reduces hallucination rates by roughly 30–70% across domains, and grounded systems can push errors below 2% on constrained tasks (6).
• Sixty-one percent of Protestant churchgoers are concerned about AI’s influence on Christianity — meaning the leader who visibly verifies will stand out in a market flooded with confident falsehood (7).
• The ninth commandment, applied to model outputs, yields one rule: verify before you testify. The founder or pastor — not the model — is the final verifier of every claim published in his name.

How Often Does AI Actually Misquote the Bible?

Start with the numbers, because they are worse than most users assume.

Bobby Gruenewald leads YouVersion, the most widely installed Bible app in history — past one billion installs, with daily engagement growing fastest in Sub-Saharan Africa (2). His team has every commercial incentive to ship an AI assistant that answers Bible questions. They have tested the frontier models against the text they know best. His verdict: the best available models misquote Scripture at least 15 percent of the time, and some models misquote it up to 60 percent of the time (1). YouVersion has therefore declined to launch a public-facing theological chatbot until accuracy and integrity can be proven — a rare and instructive act of restraint from a technology organization (2).

The failures are not merely sloppy paraphrase. Researchers and ministry technologists have documented AI systems inventing verses outright: asked for “quotes from the Bible about truth with book, chapter, and verse,” one widely used model returned a list in which verse after verse was fabricated — real references attached to words that appear nowhere in any manuscript, in any translation, in any century (3). The pattern matches what peer-reviewed studies found in other fields: in one medical study, 47 percent of AI-generated references were fabricated entirely and another 46 percent were real but inaccurate, leaving only 7 percent both authentic and accurate (8).

Even the companies building Bible study software say so plainly. Logos — whose entire business is biblical scholarship tooling — warns its own users that AI “can and will make mistakes” and must never serve as the primary source of biblical knowledge (4).

Hold that 15–60 percent figure next to another one: a third of American adults, and 34 percent of practicing Christians, now say spiritual advice from AI is as trustworthy as advice from a pastor (9). Millions of people are placing pastoral-grade trust in a system that cannot reliably quote the book it is being trusted to explain. That gap — between the trust extended and the accuracy delivered — is the defining truth-telling problem of this decade, and the church is better equipped to address it than any institution on earth.

Why Do Language Models Invent Scripture?

To respond theologically, we first need to understand the machine on its own terms — because the error is not a bug that the next update will fix. It is the system working as designed.

A large language model is, at its core, a plausibility engine. It is trained on oceans of text to predict the most likely next word, and it does this with astonishing skill. But “most likely next word” is a statistical judgment, not a truth judgment. Surveys of the hallucination literature converge on this point: models generate false, fabricated content while presenting it with full confidence and fluency, because confidence and fluency are what they were optimized to produce (5). Mitigation helps — better prompting, better models, grounding in retrieved documents — but no current technique eliminates the behavior (5)(6).

Scripture is a uniquely cruel test for such a system. The Bible exists in hundreds of translations whose phrasings circulate side by side; it is the most quoted, misquoted, paraphrased, and pseudo-quoted text in the training data. A plausibility engine swimming in that corpus will produce verse-shaped sentences with perfect biblical cadence — “The Lord does not give us more than we can handle,” “God helps those who help themselves” — that no biblical author wrote. The model is not lying, because lying requires intent. But it is also not telling the truth, because truth was never the objective function. It is doing something Scripture has a precise category for: speaking “a vision of their own heart, and not from the mouth of the LORD” (Jeremiah 23:16, ESV).

This is why hallucination is a theological category and not just an engineering ticket. Christian thinkers have begun saying this with increasing clarity: a system optimized for plausibility rather than truth demands users who are covenanted to truth (10)(11). The machine cannot care whether its words are true. Someone in the loop must. The question is whether that someone is formed for the job.

What Does the Bible Say About Truth-Telling?

Everything Scripture says about speech now applies to the speech we publish through machines. Three foundations matter most.

First: God cannot lie. Truthfulness is not one of God’s policies; it is his nature. He is “the God who never lies” (Titus 1:2), for whom “it is impossible to lie” (Hebrews 6:18). Jesus does not merely speak truth; he is the truth (John 14:6). And Scripture’s mirror image is just as stark: the devil “does not stand in the truth, because there is no truth in him… he is a liar and the father of lies” (John 8:44). In biblical theology, the line between truth and falsehood is not an information-quality issue. It is the line between two kingdoms.

Second: the ninth commandment binds the publisher, not the printing press. “You shall not bear false witness against your neighbor” (Exodus 20:16). The Reformed catechisms understood this commandment expansively — the Westminster Larger Catechism requires “the preserving and promoting of truth between man and man” and forbids “speaking untruth” in any form, careless or deliberate. Note what the commandment regulates: witness. Testimony. The act of asserting something to another person as true. When a founder publishes an AI-drafted product claim, or a pastor reads an AI-supplied “verse” to his congregation, the model has not borne witness — the human has. There is no ninth-commandment exemption for “the chatbot said it.” The catechism’s standard predates the technology and survives it untouched: if you assert it, you own it. I have argued elsewhere that this same commandment governs synthetic media and deepfakes; the misquoted verse and the fabricated video are the same sin at different resolutions (12).

Third: covenant runs on true words. The entire biblical economy — oaths, vows, contracts, witnesses at the gate, the very structure of covenant itself — assumes that words can be relied upon. “Let what you say be simply ‘Yes’ or ‘No'” (Matthew 5:37). A society where words detach from reality is a society where covenant becomes impossible, which is why the prophets treat false speech as civilizational acid: “Truth has stumbled in the public squares” (Isaiah 59:14). East African founders know this in their ledgers: in markets where formal enforcement is weak, trust is the entire currency. A business that publishes confident falsehood — even accidentally, even via a tool — is spending down the only capital that compounds.

So the theological verdict on hallucination is not “avoid AI.” It is sharper and more useful: AI’s output is unsworn testimony. It becomes witness — true or false — only when a human being adopts it and asserts it. That moment of adoption is where the ninth commandment lives now.

How Do You Use AI Without Bearing False Witness?

Here is the discipline I teach founders and pastors, in the form of a named framework you can adopt this week.

The Verified Witness Protocol

Four locks, in order. AI output passes through all four before it goes out under your name.

Lock 1 — Ground it: make the machine answer from approved knowledge. The single most effective technical control is to stop letting the model answer from its general training and start forcing it to answer from a defined body of trusted text — your actual product documentation, your actual policies, the actual text of Scripture. This approach, called retrieval-augmented generation, cuts hallucination rates by 30–70 percent across domains, and tightly grounded systems can drive errors below 2 percent on constrained tasks (6). This is not exotic technology; Kenyan builders are already shipping it — UlizaLlama answers expectant mothers’ health questions from vetted clinical content, in Swahili, by SMS, precisely because in maternal health a hallucination can kill (13). If a maternal-health nonprofit can constrain its AI to approved knowledge, your church and your company can too.

Lock 2 — Check it: verification before quotation, every time. Grounding reduces error; it does not abolish the duty to check. The rule is mechanical and non-negotiable: every quotation, statistic, name, price, and Scripture reference in AI-drafted material gets verified against the source itself before publication. For Scripture this takes thirty seconds — open the text, read the verse, confirm the words. The discipline sounds tedious until you reframe it: verification is not overhead on the work; for a people covenanted to truth, it is the work. “Verify before you testify” is simply the ninth commandment rendered as standard operating procedure.

Lock 3 — Sign it: a named human is the final verifier. Every output needs an owner — not a department, a person. The founder or a named staff member is the final verifier whose judgment, and whose name, stands behind the claim. This matters because responsibility, unlike drafting, cannot be delegated to a machine: when Scripture’s stewards give account, they give account for everything done under their authority, including by their agents (Luke 16:2). Practically: no AI-assisted content ships without a human signature in the workflow, and everyone in the organization knows whose it is.

Lock 4 — Correct it: when error escapes, confess fast. No protocol is perfect. The truth-telling organization is distinguished not by never erring but by what it does within twenty-four hours of discovering an error: public correction, plain words, no minimizing. Proverbs 28:13 is a communications policy: “Whoever conceals his transgressions will not prosper, but he who confesses and forsakes them will obtain mercy.” A visible correction culture is itself a form of witness — it tells your customers and your congregation that truth outranks image in your house.

Ground, check, sign, correct. Print it, tape it above the desk of everyone who touches a prompt.

Why Is This a Competitive Advantage, Not Just a Compliance Cost?

Here is the hopeful turn, and I mean it as more than consolation.

We are entering a market — commercial and spiritual — that will be flooded with fluent, confident, unverified text. The cost of producing plausible words has collapsed to nearly zero. Whenever the cost of producing something collapses, the value migrates to what remains scarce. What remains scarce is warranted trust: claims that have been checked, by someone with a name, who will stand behind them.

The survey data already shows the hunger. Sixty-one percent of Protestant churchgoers say they are concerned about AI’s influence on Christianity (7). Customers feel the same unease about AI-generated marketing. The leader who can say — and demonstrate — “everything we publish is verified by a human who signs it” is not performing compliance; he is occupying the most valuable position in an AI-saturated market. For the small business, this is measurable: the same verification discipline that protects your witness also protects your revenue, because AI only pays rent in an SME when its outputs can be trusted enough to act on — unverified output generates rework, refunds, and reputational debt that quietly eat the productivity gains.

And for the church, the stakes are higher and the opportunity greater. A generation is already asking chatbots the questions it used to ask pastors (9). They will encounter fabricated verses with perfect cadence. The congregation drilled in “verify before you testify” — taught to open the text and check, the way the Bereans examined even Paul’s preaching daily “to see if these things were so” (Acts 17:11) — becomes the most hallucination-resistant community in its city. Discernment, which the church has always taught as a spiritual discipline, turns out to be the digital literacy everyone else is now scrambling to invent.

YouVersion’s restraint points the way. Gruenewald did not say “never.” He said: not until accuracy and integrity are provable — and if models could quote the Bible reliably, YouVersion would help them gain access to trustworthy biblical texts (1)(2). That is the posture in full: neither prohibition nor credulity, but stewardship. Use the tool. Govern the tool. Let the God who cannot lie set the standard for everything published in his people’s name.

The fluency is the machine’s. The witness is yours. Verify before you testify.

FAQ

How often does AI misquote the Bible?
According to YouVersion CEO Bobby Gruenewald, the best AI models misquote Scripture at least 15% of the time, and some models err up to 60% of the time. Documented failures include fabricating entire verses and attaching real chapter-and-verse references to invented text, which is why YouVersion has declined to ship a theological chatbot (1)(3).

Why does AI invent Bible verses instead of saying “I don’t know”?
Large language models are optimized to predict plausible next words, not to verify truth. When asked for a verse, they generate verse-shaped text with biblical cadence whether or not it exists. Confidence and fluency are design properties, so the system states fabrications as firmly as facts (5).

Is using AI a sin if it produces false information?
Using AI is not sinful; publishing unverified claims is where the moral weight sits. The ninth commandment regulates witness — what you assert to others as true. AI output becomes your testimony the moment you adopt and share it, so the duty is verification before publication, not abstinence (12).

How can churches and businesses make AI more accurate?
Ground the system in approved knowledge — your verified documents, policies, and the actual biblical text — using retrieval-augmented generation, which cuts hallucination rates by roughly 30–70%. Then add human checks: verify every quotation against its source, and require a named person to sign off before anything publishes (6)(13).

Should Christians stop using AI for Bible study?
No — but AI should never be the primary source, a warning even Bible software maker Logos gives its own users. Treat AI as a study assistant whose every reference you confirm in the open text, the way the Bereans tested Paul’s preaching against Scripture daily (4).

Related Reading

• UlizaLlama and Grounded AI: Making the Machine Answer from Approved Knowledge
• Does AI Actually Pay Rent? The Evidence from Small Businesses
• A Third of Christians Trust AI as Much as Their Pastor
• Deepfakes and the Ninth Commandment: False Witness at Machine Scale

Sources and Evidence

1. Christian Post — “AI’s Bible misquotes range from 15% to 60%: YouVersion CEO” — Direct reporting of Bobby Gruenewald’s stated misquote rates; primary-source interview coverage of the executive responsible for the world’s largest Bible app.
2. Christian Post — “YouVersion founder talks concerns about pastors’ embrace of AI; app reaching 1B” — Gruenewald on YouVersion’s decision not to ship AI answers until accuracy is provable; corroborated by Christian Daily International.
3. Missio Nexus — “AI Hallucinations, Chatbots, and the Truth of Holy Scripture” — Documented cases of AI fabricating Bible verses with real references attached; mission-sector technology analysis.
4. Logos — “AI and Bible Study” — A commercial Bible-software maker warning its own users that AI “can and will make mistakes”; notable as testimony against interest.
5. Frontiers in Artificial Intelligence — “Survey and analysis of hallucinations in large language models” — Peer-reviewed survey establishing hallucination as a persistent design property rather than an eliminable bug.
6. SQ Magazine — “LLM Hallucination Rate: 40+ Stats (2026)” — Aggregated statistics on retrieval-augmented generation reducing hallucinations 30–70%, with grounded systems below 2% on constrained tasks; secondary aggregator, figures consistent with the peer-reviewed RAG literature (example).
7. Lifeway Research — “Pastors, Churchgoers See AI as Concerning and Confusing” (Sept 2025 survey) — Probability-sampled survey of 1,003 Protestant pastors and 1,200 churchgoers; source of the 61% concern figure.
8. Cureus / NIH — Study on authenticity and accuracy of ChatGPT-generated references — Peer-reviewed finding that 47% of AI-generated medical references were fabricated and only 7% fully accurate; establishes the cross-domain pattern.
9. Barna Group — “AI is Becoming a Spiritual Authority, Even Among Practicing Christians” — Barna/Gloo State of the Church research; source of the one-in-three trust statistic.
10. Aleteia — “AI’s Impact on Truth and a Christian Response” — Christian ethical framing of plausibility-optimized systems.
11. Christ Over All — “A Christian’s Perspective on Artificial Intelligence” — Long-form theological treatment of AI and truthfulness from a Reformed evangelical perspective.
12. Orthodox Presbyterian Church, New Horizons — deepfakes as a ninth-commandment issue — Confessional Reformed application of the false-witness prohibition to synthetic media.
13. Jacaranda Health — “Jacaranda launches first-in-kind Swahili Large Language Model” (UlizaLlama) — Primary source on a grounded, domain-constrained African LLM deployed in maternal health.