Can you trust what AI tells you about the Bible?

Many people now ask AI about faith

Questions about faith now reach chatbots in very large numbers. OpenAI reported about 70 million religion-related requests to ChatGPT each week, from an analysis of requests made between September 2025 and September 2026; 34 percent concerned Christianity, and about 22 million a week involved studying religious texts [9]. A sizable minority of Americans already use AI for spiritual questions. An NPR/Ipsos poll of 1,011 US adults, taken from 7 to 9 August 2026, found that 61 percent had used an AI tool for some purpose [1]. Among those users, 15 percent had used AI to study Scripture, 13 percent to write prayers, and 11 percent to seek spiritual guidance [1]. The same poll found that nearly seven in ten AI users said they would never turn to AI for religious or spiritual guidance, so the picture is mixed [1].

Trust is a separate question from use. In a survey of 1,514 US adults conducted online in November 2025, Barna Group, working with Gloo, found that 30 percent agreed that spiritual advice from AI is as trustworthy as advice from a pastor [2]. Agreement was 34 percent among practicing Christians, 39 percent among Gen Z, and 44 percent among millennials [2]. The same research recorded real worry. Some 74 percent of US adults, 83 percent of practicing Christians, and 94 percent of Protestant pastors said they were concerned that AI could misinterpret Scripture [2].

Researchers have started to test the answers

Independent testing shows that chatbots often leave faith out of answers where people expect to find it. In 2026 a consortium led by Brigham Young University, together with Baylor University, the University of Notre Dame, and Yeshiva University, released the AllFaith Benchmark, a set of hundreds of real ethical questions drawn partly from ChatGPT transcripts and partly from contributors in faith communities [4]. The group, called the Consortium for Evaluation of Faith and Ethics in AI (CEFE-AI), reported launch results for 14 large language models and surveyed 1,125 Americans about what they expected to see in an answer [4]. The team also reported that only 0.2 percent of more than 12,000 research papers on bias in AI had addressed religious bias at all [4].

Axios reported the central result. People expected religion to appear in answers from 45 to 59 percent of the time, depending on the topic, while the models mentioned it from 5 to 16 percent of the time [3]. On grief and loss, for example, 59 percent of respondents expected a religious perspective, and the models offered one in 16 percent of their answers [3]. The researchers also reported that every model they tested showed a repeatable pattern of treating some traditions more favorably than others [3].

The Gospel Coalition's Keller Center for Cultural Apologetics ran a different kind of test. In its AI Christian Benchmark, seven Christian scholars graded how seven AI platforms answered seven frequently searched questions about the Christian faith, such as whether Jesus rose from the dead [5]. Scores out of 100 ran from 40 for the lowest model to 64 for the two highest [5]. The authors attributed much of the spread to how each company tunes its models after training, which is their interpretation rather than a measured result [5]. This benchmark graded answers against historic Christian teaching as its evangelical Protestant reviewers understood it, a yardstick that Catholic, Orthodox, or other reviewers might set differently.

How a chatbot produces an answer

Unless it is connected to search or to a set of checked sources, a large language model does not look up an answer the way a reader looks up a verse; it generates the text that is statistically likely to come next. Some chatbots now search the web or a library before they answer, which makes invented text less likely but leaves the quality of the sources and of the summary to be checked. OpenAI explained in September 2025 that its models first learn by predicting the next word across very large amounts of text, and that this training text carries no labels marking statements as true or false [6]. Patterns that recur constantly, such as spelling, become reliable as models grow. Rare and specific facts are harder, and a model can produce a fluent sentence that is simply wrong [6]. OpenAI added that common training and testing methods "reward guessing over acknowledging uncertainty" [6].

For questions about the Bible this matters in two ways. First, a chatbot does not by default show where its wording came from, so a reader cannot tell whether a sentence is a quotation, a paraphrase, or a blend of several sources. Second, references can be invented. A 2023 study in Scientific Reports checked bibliographic citations produced by ChatGPT and found that 55 percent of those from GPT-3.5 and 18 percent of those from GPT-4 pointed to works that do not exist [7]. That study tested academic references from 2023 models, not Scripture or today's systems, so it shows the kind of error that can occur rather than how often it occurs now. A confident tone is not evidence.

Practical checks any reader can apply

The simplest safeguard is an old one: read the passage yourself. The book of Acts commends the people of Berea because they "searched the scriptures daily, whether those things were so" (Acts 17:11, KJV) [8]. A few habits bring that practice to AI answers.

Ask for the reference. If a chatbot says the Bible teaches something, ask which book, chapter, and verse it has in mind. Then open a Bible, in print or online, and read the verse along with the paragraph around it. A real verse lifted out of its setting can still mislead, and a reference that leads nowhere tells you the answer was built on sand.

Compare translations. Reading the same verse in two or three translations shows where the wording is settled and where translators made different choices. Older public-domain translations, such as the King James Version, can be read freely online beside modern ones. If an AI quotes a verse in wording that no published translation uses, treat the quotation as unverified until you find it.

Ask which tradition holds the view. On a number of questions, Catholic, Orthodox, and Protestant churches teach differently, and each sets out its teaching in its own published documents. A useful answer names whose teaching it describes. If a chatbot presents one position as simply what Christians believe, ask what the other traditions say, and then check each account against that tradition's own sources.

Check against a published commentary. Commentaries, study Bibles, and catechisms are written by named authors and pass through editors, so their reasoning can be followed back to its evidence. When an AI answer and a respected commentary disagree, the commentary is the better place to start, because it shows its work.

Where AI can still help

Used with care, a chatbot can be a starting point rather than an authority. It can suggest passages to read, lay out the main positions in a debate, or explain an unfamiliar term, and each of those can then be checked against a Bible and a reference work. The surveys suggest that many people already lean on it more heavily than that. In the Barna research, 48 percent of practicing Christians said they trusted AI to help them grow spiritually, compared with 12 percent of Protestant pastors [2]. Whichever group a reader belongs to, the gap between those figures is a reason to keep the habit of checking.

What this means for the reader

You can trust an AI answer about the Bible as far as you can check it. Ask for the verse, read it in context, and see what each tradition and a good commentary say before you rely on what the chatbot told you.