Scripture Quotation in Generative AI

Long after the apostles breathed their last, one of the Church’s enduring tasks remains: to pass down God’s unchanging word without letting it be lost, corrupted, or silenced. Few responsibilities are more sacred than this act of preservation and transmission. That task has never been static. Every generation has faced new tools and new temptations to alter the text it receives. Used without care, generative AI has the capacity to fabricate, misquote, or subtly distort Scripture at a scale no scribe could match. But the origin of its error lies with us, either directly or indirectly, based on the decisions we make as we implement it. The faithful must therefore meet this challenge head-on if we are to hand down God’s word to this generation intact.

The purpose of this paper is four-fold. First, we will briefly survey some of the historical challenges for textual transmission of Scripture. Second, we will make the case that Christian technologists must not shirk their responsibility to preserve Scriptural fidelity in spite of the challenges that generative AI presents. Third, we will present a study demonstrating that an output transform layer sourced from a Bible API can quote Scripture with perfect fidelity, and we will show why the alternatives fall short. Finally, against the backdrop of these results, we will conclude with an open appeal to Bible translation licensors to loosen their restrictions on the use of their translations in generative AI.

The grass withers, the flower fades; but the word of our God stands forever.
Isaiah 40:8 (Web English Bible)

AI is not the first technology to expose pressures in the transmission of Scripture. In 1516, racing to publish before a rival Greek New Testament could appear, Erasmus issued the first published Greek New Testament after preparing it in only a matter of months. The one manuscript of Revelation available to him lacked its final leaf, including Revelation 22:16–21. To complete the text, Erasmus translated the missing verses back into Greek from the Latin Vulgate, creating several readings unattested in the Greek manuscript tradition known to him. Through later editions in the Textus Receptus tradition, some of these readings influenced the King James Version and remained influential in English Protestant Bible use for centuries.^[7]

In spite of such cautionary tales, it’s important to understand that the printing press itself did not corrupt Scripture; rather, haste, a limited manuscript base, and inadequate verification allowed particular editorial errors to be reproduced widely. And it was that same technology which exacerbated the error that also made later correction and comparison possible. The printing press is almost unanimously heralded as a giant leap forward for Christendom. Gutenberg’s press made Scripture affordable. A scribal Bible cost months of labor and a small fortune; within fifty years of 1455, printed Bibles circulated by the hundreds of thousands. Erasmus’s Greek New Testament (1516), Luther’s German (1522), and Tyndale’s English (1526) put the text directly in scholars’ and laypeople’s hands faster than authorities could suppress it.

Sharing God’s word has never been without risk; just ask the martyrs of the persecuted church when you meet them in heaven. But mistakes in Scripture transmission happen much more frequently than we’d care to admit. We struggle to remember that one Bible verse and get it slightly wrong; the pastor accidentally skips a word during his sermon’s Scripture reading; a medieval artist paints horns on Moses due to a translation ambiguity.^[10] These are honest mistakes, to be sure — but technically mishandlings of Scripture nonetheless.

A Christianity Today article in 2005 estimated that churches systematically send well in excess of 2 million youth on short-term mission trips every year from the U.S. alone.^[11] It is inconceivable that the majority of these kids possess the spiritual maturity, Scriptural foundation, and relational empathy to not pose a significant risk of misrepresenting Christ and God’s word in cross-cultural contexts — but we continue to send them nonetheless. Whether or not we should do so is the subject of a different paper the author is not qualified to write. The point is, however, that the Church has grown accustomed to and celebrates systematic risk taking in pursuit of fulfilling the Great Commission. Yet it often has a clear double standard when it comes to the risk of deploying new technologies and places the bar unreasonably high in comparison to its other endeavors.

AI models have become almost miraculously powerful, and we must respect that power. Given a purpose, a task, or a goal, AI models have displayed remarkable ingenuity and tenacity in achieving it. If we fail to wisely instruct them or carelessly eschew safeguards, any fault or undesirable consequence inevitably lies with us. Its probabilistic nature is the source of its power, not something to fear. We are accountable to set its boundaries in order to not let it lead others astray. Yet Christ exhorts us to risk much for the Kingdom — not recklessly, but with sober discernment — for with great risk there is also great reward.^[12]

AI models have become almost miraculously powerful, and we must respect that power.

Modern agentic AI systems consist of much more than just the model. In fact, given the convergence of model capabilities, the industry now widely regards the harness to have an equal or greater impact on outcomes than the model itself. Don’t misunderstand the contention as the model being unimportant in an absolute sense. Rather, the argument is that anyof the most recent generation models are well capable of achieving the desired result given the proper context, and it therefore becomes less important which model in a given class that is chosen, and far more important what the scaffolding around the model is. The models themselves have largely become commoditized. While there may certainly be differences between classes of models, there is not much relative distance between the capabilities of a model by one lab vs another in the same model class.

Context is king. Even a low powered model that is handed relevant proprietary context can do a serviceable job of using that context to accurately respond. But even the most powerful model can do very little without the proprietary context it requires. Therefore, in order to fully examine possible failure modes in Scripture quotation, we must enumerate the various components and processes at play in a typical agentic harness.

The Model

Synthetic reasoning capabilities are owned by the model. It is the underlying machinery that formulates a response. While the quality of its response is heavily dependent on the tools and context handed to it, the model itself must use that context and those capabilities properly for them to unlock the desired outcome. The model is trained on vast amounts of data, generally including Scripture, biblical commentaries, and works authored by Christian authors. Therefore, most models have a functional understanding of these matters from their training in at least some capacity.

In the lifecycle of a text-based generative AI request involving Scripture, the model is at a minimum responsible to derive which Bible verses to reference. This necessarily requires a certain level of basic Scripture interpretation, depending on one’s definition of the word. For example, the model must be able to derive that Genesis 11:7 pertains to the story of the Tower of Babel based on the surrounding context in order to correlate it with a request for that context. Strictly speaking, this is a kind of interpretation.

Guardrail Gates

Guardrails in a general sense are any safety mechanism that prevents a suboptimal outcome. In the context of an agentic harness, a discrete guardrail may actually gate processing or the final output of a response. One example already given is OpenAI’s guardrail which refuses to reproduce more than 25 words of copyrighted material. Guardrails may also be used to evaluate a generated response and short-circuit the response with a deterministic message, or send an enhanced input to a model to retry the initial request.

System and User Supplied Instructions

Models have default operating instructions, but generally are given more specific instructions by the caller. General behavior requirements, how to treat Scripture, and any other instruction pertinent to the successful quotation of Scripture may all be supplied. In addition, the user’s own instructions shape the goal of the model so long as they do not countermand the model’s system instructions.

Queried Context

Retrieval Augmented Generation (RAG) is the general term for a harness querying an external system for relevant content, and then subsequently handing the model that context by injecting it into the supplied operating instructions. This turns out to be one of the most powerful mechanisms at our disposal to shape the response of the model. By constraining the knowledge that is retrieved to only the sources we trust, we can effectively ground the model in our preferred worldview. Instead of relying on the base model’s knowledge and interpretation of Scripture, we can provide relevant context with which to derive its answer. The sequence of operations and deterministic mechanism are what differentiates this concept from the next component we will explore.

Tool Calls

Recent generation models are able to call out to defined external services as part of what’s called the agentic loop. Whereas RAG queries an external source and injects context prior to handing the request to the model, tool calls are performed while the model is in the process of generating its response. This is an extremely important distinction. Tools (external services) are made available to the model to use at its own discretion over the course of up to a preconfigured number of turns. The model may or may not choose to utilize a tool it’s given as the model’s probabilistic properties are at play.

Output Transforms

Once a response is returned by a model, the harness may have one or more deterministic processes that alter the final output. For example, a routine could be programmed to identify any URLs and automatically convert them into proper hyperlinks.

Any of the most recent generation models are well capable of achieving the desired result given the proper context, and it therefore becomes less important which model in a given class that is chosen, and far more important what the scaffolding around the model is.

Design

Purpose and Scope

The purpose of this study is to demonstrate the most effective Scripture quotation method for use in conversational AI given a predetermined Scripture reference. The intention is not to advocate for or provide insight into an AI model preference, Bible translation, or model temperature setting. These variables have been introduced solely as controls to validate that the recommended Scripture quotation mechanism works across these variables, thereby isolating the efficacy of the methods themselves.

Further, this study does not speak to a particular model’s ability to select a correct Scripture reference, nor does it make any claims about any particular model’s ability to correctly interpret Scripture. These are matters heavily influenced by and enmeshed in a given agentic system’s overarching harness and cannot be easily disentangled from it. We’ve had enormous success on our platform through curating an extensive corpus of trusted sources from biblically grounded authors. This has greatly reduced the failure rate in terms of Scripture selection because we’re able to inject the human authored content to do the heavy lifting of Scripture interpretation and not leave the model to make such weighty decisions. However, the test harness used in this study is completely isolated and has no integration with our platform. The results of this study have implications only for determining the best method for precisely quoting Scripture in agentic AI systems given a pre-selected passage to quote.

Scale

In total, 33,000 observations were made in the test run. 3 iterations were run for each sample permutation to account for any ephemeral issues and probabilistic model variance in response. Up to 5 retries per sample permutation were allowed for ephemeral network errors to ensure a timeout or temporary API outage didn’t present in the data as misquotation. If after 5 retries any external service still failed, that sample was discarded from consideration and did not impact the averages presented in the results. 4/10 of the selected models do not support the temperature setting; therefore to simplify the math, we represent the temperature setting as 2.2 rather than 3 in the formula to derive the total number of observations.

{
 [ (6 models x 3 temperatures) + (4 models x 1 temperature) ]
 / [10 models x 3 temperatures]
}
x 3
= 2.2

5 methods
x 10 Bible translations
x 10 Scripture passages
x 10 models
x 2.2 temperature settings
x 3 iterations
= 33,000

AI Models

10 representative LLMs were chosen for the study. They range from low to high power, both commercial and open source.

Bible Translations

10 representative English Bible translations were chosen for the study. They include examples from all 4 supported Bible API providers, both public domain and commercially licensed.

Scripture References

10 representative Scripture references of varying length and obscurity were chosen for the study, spanning both Old and New Testaments.

Model Temperatures

6 of the 10 models tested support a temperature parameter. Temperature dictates how much probabilistic liberty will be granted to the model in deriving its response. Realistically, any application responsible for Scripture quotation should set this parameter within the 0.0 to 0.5 range. 3 temperature setting variations were used: 0.0, 0.25, and 0.5.

Scripture Quotation Mechanisms

5 Scripture quotation mechanisms were tested in this study; they are enumerated in the section above.

Results

The research study reveals that the output transform method has a perfect 100% success rate of faithfully quoting known Scripture references in tested English translations. Quotation of 10 non-English language Bible translations deviated only slightly at 99.4%. For the output transform method, we instruct the model to simply leave a placeholder Scripture citation which the harness then deterministically replaces with the response from a given Bible API for the cited Scripture passage.

Not only is the output transform method superior in accuracy, but its failure modes also result in less consequential outcomes than other methods. The impact to the end user experience is merely a Scripture reference placeholder that is not successfully replaced with the corresponding full Scripture quote. In other words, even an exceedingly rare failure of this method does not result in a misquotation of Scripture whatsoever. Moreover, the results hold uniformly against all tested models, model temperature settings, Bible translations, and Scripture references.

Aggregate results are reproduced below, and the full result set is available here. We have also open sourced the test harness that produced these results for the sake of transparency and as a gift to our community of fellow Christian builders. You may re-run these tests with any combination of methods, models, Bible translations, Scripture references, and model temperature settings. The repository was developed by Anthropic’s Fable 5 model, the most powerful and capable commercially available model on the market at the time of writing. Care was taken to present the model with an impartial description of each method so as not to skew the implementation to favor one method over another. All prompt templates derived to test the various methods were entirely constructed by the model based on a description of the study’s goals. There have been several enhancements since, which are all available in the repository’s commit history for auditing purposes.

In the tables below, the “Exact” column is a measurement of comparison inclusive of punctuation and all other characters other than whitespace. The “Normalized” column is a measurement of comparison with all punctuation and letter casing distinctions discarded.

Scripture Quotation Method

The output transform method is the clear winner for exact reproduction of a Scripture passage, though the top spot is tightly contested for English translations when Scripture is normalized (punctuation and letter casing not considered). This is likely due to the model stripping out punctuation and verse markers it feels aren’t relevant for RAG and tool calls.

Every decision made by a probabilistic AI model is an opportunity for variation. In a conversational AI context, that’s normally a feature rather than a bug. After all, users want to have nuanced, contextualized conversations. But that infrastructure is catastrophic when it comes to faithfully quoting Scripture. This is a case where we explicitly don’t want variation or creativity; we want God’s word, faithfully conveyed in its perfection.

Therefore, the most successful Scripture quotation method minimizes probabilistic decision points and maximizes deterministic processes. We ask the model to do only what it must, and delegate everything we can to the harness. Assuming input and output guardrails, the only other decision points we are deferring to the model is the Scripture selection and reference generation. In other words, the model just needs to know what verses to use and how to reference them, and nothing more.

Agent tool calls retrieved the exact text to quote, but the model still failed to quote the passage with correct punctuation and verse markers 3.8% of the time. We observed that even the most capable models failed to always call a provided tool when quoting Scripture, presumably under the assumption that invoking the tool for a reference it thought it knew based on its training was inefficient, and therefore it would be optimal to skip what it considered an extraneous tool call (see 1 Corinthians 8:2). Even so, the tool calls method may be considered “good enough” by many (especially for English translations), and indeed seems to be the more straight-forward architecture. If you don’t have enough control of your pipeline to use the output transform method but do have the ability to add tool calls to the agentic loop, this method is a reasonable backup.

The simulated RAG method for this study hands the model the exact Scripture to quote on a silver platter. And yet 7% of the time, the model received the exact text to quote, and then still immediately turned around and failed to reproduce it accurately. And remember, this represents a ceiling as the simulated RAG method bypasses the actual retrieval mechanism by assuming 100% success rate from the RAG pipeline. There are still several things that could go wrong in a real world scenario, including parsing the Bible reference, sending the request, and parsing the response. The RAG method can be effective assuming a highly deterministic pipeline that perfectly references the intended Bible passage and perfectly queries for the intended Bible passage, which is a bar that is not necessarily straight-forward to achieve.

The unbounded web search and unassisted model methods may safely be dismissed out of hand; the web search is superior to the unassisted model but only just, and it does not even crack 65% quotation accuracy when normalized. It would be wise to stay away from both when Scripture quotation is important to your application. While some models fare relatively better than others, all of them that we have tested are objectively atrocious at faithfully quoting Scripture. As discussed, this is not surprising given the underlying probabilistic architecture of LLMs. When Scripture quotation fidelity is important, you rely on the raw model unassisted at your own peril.

AI Models

Average successful Scripture quotation clustered within 22.3 percentage points depending on the model, from 59.3-81.6%. Grok 4.5 was a notable outlier at nearly 4 percentage points out in front of the second highest score, while Claude Opus 5 interestingly gets the notoriety of being the worst performer at 59.3%, though it scored a more respectable 82.7% for its normalized score. Proverbs 3:5-7, anyone?

Nonetheless, there seems to be no strong correlation with a model’s class and its ability to accurately quote Scripture. If anything, lower-powered models scored higher on average. It seems being too clever may actually be a handicap when it comes to faithfully quoting Scripture (see Proverbs 26:12). However, any hypothetical guidance on model selection is immediately rendered moot once the model has any other mechanisms at its disposal. As we’ve seen, it’s the harness that ultimately moves the needle.

Scripture Passages

The particular Scripture passage quoted has a significant impact on the fidelity of the quotation. Genesis 1:1 was the clear winner. Predictably, the length of the passage had a sizable adverse effect on whether the passage was quoted 100% accurately. It was somewhat surprising, however, that John 3:16 did not score higher. And of course we would be remiss not to bemoan the poor showing from 1 Peter 3 given our affinity with apologetics.

Model Temperature

The model temperature parameter seems to have had a negligible effect on accurate Scripture quotation (2.5 percentage points — 72.7-75.2% — with temperature values ranging from 0.0 to 0.5). Somewhat counterintuitively, it appears that lower temperature actually has an inverse correlation to the model’s ability to accurately quote Scripture. This would indeed be a notable finding were the variance much greater. As it stands, however, the difference is too small to be considered actionable. The key takeaway is simply that temperature, within reason, does not have a significantly adverse effect on Scripture quotation fidelity.

Bible Translations

Two primary factors seem to influence how well a Bible translation can be quoted by a given model unassisted: its popularity and its accessibility. The New International Version leads the pack, presumably due to its historical popularity in the past few decades and widespread use online. However, translations like the Berean Study Bible also rank highly through better representation due to being public domain and freely available online.

Languages

A subsequent run of the test harness was made against 10 non-English Bible translations, 1 from each of the 10 most spoken non-English languages. Non-English Bible translations were misquoted on average 12.96 percentage points more often than English translations.

However, that degradation was not evenly distributed amongst methods. The table below compares non-English vs English exact and normalized averages. In languages other than English, the exact quotation dropped only 0.6pp for the output transform method and 0.8pp for the simulated RAG method ceiling. But the disparity began to grow sharply the more the model had to make decisions in non-English languages. Tool call exact quotation accuracy dropped by 10.8pp, web search by 26.5pp, and the unassisted model by 25.3pp. Therefore, while the tool call method is a reasonable 2nd choice for English translations, its efficacy drops precipitously for non-English translations. The output transform method is in a class all its own for non-English Bible translations.

The research study reveals that the output transform method has a perfect 100% success rate of faithfully quoting known Scripture references in tested English translations.

1. Alan R. Millard, “‘Scriptio Continua’ in Early Hebrew: Ancient Practice or Modern Surmise?” Journal of Semitic Studies 15, no. 1 (1970): 2–15.
2. Philip W. Comfort, New Testament Text and Translation Commentary (Carol Stream, IL: Tyndale House, 2008), 663.
3. Bruce M. Metzger and Bart D. Ehrman, The Text of the New Testament: Its Transmission, Corruption, and Restoration, 4th ed. (New York: Oxford University Press, 2005), 25.
4. Emanuel Tov, Textual Criticism of the Hebrew Bible, 3rd ed., rev. and exp. (Minneapolis: Fortress, 2012), 54–59.
5. Peter M. Head, “A Case against the Longer Ending of Mark,” Text & Canon Institute, June 14, 2022, https://textandcanon.org/a-case-against-the-longer-ending-of-mark/.
6. David Norton, A Textual History of the King James Bible (Cambridge: Cambridge University Press, 2005), 153–55.
7. F. F. Bruce, History of the Bible in English: From the Earliest Versions, 3rd ed. (New York: Oxford University Press, 1978), 23.
8. Jacques Ellul, Propaganda: The Formation of Men’s Attitudes, trans. Konrad Kellen and Jean Lerner (New York: Alfred A. Knopf, 1965; repr., New York: Vintage Books, 1973), 121.
9. Horace He and Thinking Machines Lab, “Defeating Nondeterminism in LLM Inference,” Thinking Machines Lab: Connectionism, September 10, 2025, https://doi.org/10.64434/tml.20250910.
10. William H. C. Propp, “Did Moses Have Horns?” Bible Review 4, no. 1 (1988): 30–37, 44.
11. Kurt Ver Beek and Robert Priest, “Are Short-Term Missions Good Stewardship?” Christianity Today, July 30, 2005, https://www.christianitytoday.com/2005/07/are-short-term-missions-good-stewardship/.
12. Ted Halum and Jake Carlson, “Generative AI in Christian Evangelism,” The Apologist Project, 2025, http://apologistproject.org/generative-ai-in-christian-evangelism.
13. Vincent Matinde, “AI’s Scripture Problem: Misquotes Range from 15% to 60%, Says YouVersion CEO,” Christian Daily International, March 16, 2026, https://www.christiandaily.com/news/ais-scripture-problem-misquotes-range-from-15-to-60-says-youversion-ceo.
14. Benjamin Kaiser, “Can LLMs Accurately Recall the Bible?” benkaiser.dev, 2024, https://benkaiser.dev/can-llms-accurately-recall-the-bible/.
15. Vadym Arnaut, “Don’t Trust the LLM with Scripture: A Canonical-Text Substitution Layer for Bible Quotes,” Dev.to, May 10, 2026, https://dev.to/arvavit/dont-trust-the-llm-with-scripture-a-canonical-text-substitution-layer-for-bible-quotes-17cg.
16. Alex Chao, “When Not to Generate: How AI Systems Quote Scripture, and What Authoritative Quotation Requires,” Fide AI, 2026, https://github.com/FideAI/scripture-quotation-fidelity/blob/main/papers/p01-scripture-quotation/paper/main.pdf.