Transparency
How the study assistant works
Which model writes it, what it is given and what it is never given, which tradition it reads from, how its citations are checked against the text, and who is answerable for the answers. Written to be read before you decide how much weight to put on anything it says.
What it is
A general-purpose language model, prompted for this job
Nothing here is a Bible-specific model, and nothing has been trained or fine-tuned on Scripture by this project. The assistant is a commercial language model given a written instruction, the passage you are reading, and — for a word study — the lexicon entry for the word. Its skill at Greek and Hebrew is the model’s own, which is exactly why the two safeguards further down this page exist.
This deployment currently uses a different model depending on the action:
- Summarize —
google/gemini-3.5-flash-lite - Key themes —
anthropic/claude-sonnet-5 - Insights —
anthropic/claude-sonnet-5 - Word study —
anthropic/claude-sonnet-5
The administrator of this instance can change these, so the list is read live rather than written down here.
What it is given
The text of the passage, chapter or book you asked about; the reference itself; the word you asked about and, when you reached it by tapping a tagged word, that word’s lexicon entry; and, only if you tick the box, the results of a web search. The written instruction also carries any of the guiding principles listed under reader feedback below that apply to the passage and study type. That is the whole of it.
Your own material is never sent. Not your notes, not your highlights, not your bookmarks, not the documents in your library, not your profile, and not anything you have written anywhere else in the app. Each request stands alone: the assistant has no memory of your previous questions and is told nothing about you.
In a licensed translation the passage is re-read from the public-domain Berean Standard Bible first, because a publisher’s wording is not ours to send to an AI provider. The answer says when this has happened. See the copyright page.
Every verse it cites is checked
Misquoted Scripture is the best-documented failure of AI Bible tools. YouVersion’s founder has put the rate at “at least 15% of the time” for the best models on the most-indexed translations, “some of them as much as 60%”, and it is why they have declined to ship a public question- and-answer bot at all. The failure is a specific one: a reference is attached to a claim, the claim reads well, and the reference is to a different verse or to no verse.
This app has something a chat product does not — the text itself. So when an answer finishes, every reference in it is looked up in the same copy of the Bible you read from, and the result is shown under the answer with the words that are actually there. No model checks another model’s work: the references are read out of the prose and the verses are read out of the database.
What that claim covers, exactly. A citation “checks out” when the reference exists and reads as printed. It is not a claim that the verse supports the sentence it was attached to — no automatic check can make that claim, which is why the words are put in front of you instead. It also verifies references rather than quotations: a loose paraphrase in the prose with a correct reference beside it will pass.
It reads from a tradition, and says which
In January 2026 the Bible Society, with the Cambridge Centre for Chinese Theology, tested five Bible chatbots and found an overwhelming lean towards US evangelical readings, presented as though they were simply what the text says. The problem identified was not that the answers had a viewpoint — every reading of Scripture has one — but that readers were never told, and so had no reason to think another reading existed.
So the tradition is a setting, and it is named at the top of every answer it produced. The 11 choices are:
- Across traditionsdefault
- Reformed
- Lutheran
- Anglican
- Baptist
- Methodist / Wesleyan
- Pentecostal / Charismatic
- Anabaptist
- Roman Catholic
- Eastern Orthodox
- Historical-critical
The default reads across traditions, and it is labelled on answers like any of the others — because setting out both sides of a disputed passage and leaving it standing treats as open what some traditions consider settled. That is a position, not the absence of one, and hiding it behind a blank label would be a smaller version of the same problem.
A reader can also set their tradition beside others. In the assistant, up to 3 further traditions can be chosen for a particular question, and the answer is then written once from inside each of them in turn, under a heading naming it, closing with what those traditions actually disagree about in that passage — or with a plain statement that they do not, since most of Scripture is not disputed between them. The reader’s own tradition is always one side of it and cannot be dropped: the comparison exists to put the reading they are being given next to another, and the model is asked not to close by recommending one of them.
It learns from reader feedback — in the open
Every answer carries a thumbs control
Under every finished answer you can say whether it helped, and say why. Each report goes to a human review queue with a copy of the answer and a record of which passage, study type, tradition and model produced it — a person reads every one, and a wrong answer that other readers were still being served from the cache can be withdrawn on the spot.
Nothing you write is ever pasted into the prompt. A report that turns out to be right is distilled, by the administrator, into a short written principle that is added to the assistant’s instruction for the studies it applies to. That distinction is a safety line, not bureaucracy: an instruction any account could write into would be a way for one reader to steer every other reader’s answers.
The current principles, read live
There are none yet: no reader report has been turned into standing guidance so far, and the instruction is exactly the one described on this page. When one is added it will be listed here, in full, with where it applies.
What it is not for
It is not pastoral or doctrinal authority
It is useful for orientation and for finding threads worth pulling. It is not a minister, a counsellor, or a substitute for your own reading, and it should not settle anything that matters to you on its own. Every answer says as much at the foot of it.
It is not a crisis service, and nobody is watching
Nothing you write in this app reaches a person in real time. If what you type suggests you may be in trouble, the app puts these numbers in front of you — it does not block anything, and it records, reports and sends nothing, including to whoever runs this instance.
If you are in immediate danger, contact your local emergency services. Otherwise these lines are free, confidential and open around the clock:
- United States and Canada: 988 Suicide & Crisis Lifeline — Call or text 988
- United Kingdom and Ireland: Samaritans — Call 116 123, free, any time
- Anywhere else: Find a Helpline — A free, confidential line in your own country
Who is accountable
There is no theological advisory board
Independent reviewers of this category have noted that none of the leading AI Bible apps names one. This app is currently no exception, and saying so is more use to you than an impressive-sounding page would be. In practice that means: the instructions the model is given were written by the administrator of this project, not reviewed by clergy or by academic theologians; no denomination or institution has endorsed anything here; and the tradition descriptions above were written from general knowledge rather than checked with each tradition.
Scripture Study is run by one administrator, who is responsible for all of it. If that changes — if anyone reviews the prompts and agrees to be named — they will be listed here, alongside exactly what they reviewed rather than a general endorsement.
Telling us an answer was wrong
This is the useful thing you can do with a bad answer, and it is worth doing even when the mistake looks obvious — the prompts are changed on the strength of reports like the one you'd send. The fastest route is the thumbs under the answer itself (see above), which carries the answer and its context with it; these routes work too:
If you have an account, the quickest route is Send feedback in the sidebar. It reaches the administrator with your account attached, so there is nothing to prove before your request can be acted on.
If you do not — you asked for access and never got one, you are a publisher, or you are writing on someone else’s behalf — use the contact page. It goes straight to the administrator, and the reply comes back to the address you give.
For how to use the assistant rather than how it is built, see the help article. For what is stored about your use of it, see the privacy notice.