---
title: "AI in These Studies"
category: "other"
description: "How AI is used to research and draft the studies on this site: what it does, the failure modes it introduces, the guardrails against them, and where its usefulness ends."
tags: ["ai", "study-method", "transparency"]
draft: false
date_created: 2026-07-25
date_modified: 2026-10-22
ai_provider_models:
  - anthropic/claude-opus-5
  - anthropic/claude-opus-5.5
  - anthropic/claude-sonnet-5.5
---

# AI in these studies

AI assists the research and drafting of the studies on this site. This page sets out the method: what
the tool does, the failure modes it brings with it, and where its usefulness ends.

!!! quote "The short version"

    AI is a **research assistant**, not an author and not an authority. It reads the languages I
    can't yet read, compares translations I can't hold in my head at once, and finds the case
    against an idea faster than I can. Every claim resolves to a source, and a human reads every
    line against Scripture before it is published.

Want the same setup for your own study? [Getting Started](https://the-way.lewy.au/resources/getting-started/) walks through it.

---

## What it does

**The original languages come within reach.** I am not formally trained in Biblical Hebrew, Aramaic,
or Greek — I'm at the start of that road, not the end of it. Working through a passage's morphology,
lemmas, Strong's numbers, and semantic domains by hand would take days per chapter, if I could do it
reliably at all. Querying a structured dataset takes minutes, and the result is checkable: when a
study says a word is <span dir="rtl">חֶסֶד</span> (*chesed*, H2617) and gives its range of meaning, that
came from a lexical database with an entry anyone can look up.

**Many translations at once.** A single English translation is one committee's set of decisions.
Comparing the Masoretic Text and Septuagint against several English renderings surfaces the places
where translators disagree — which is where a study should slow down and say so.

**Thoughts get structured fast.** A half-formed idea becomes an outline, a set of questions, and a
list of passages to test it against, in an evening rather than over months. Speed here isn't about
producing more; it's about being able to *finish* the thought instead of abandoning it.

**Assumptions get disproved.** This is the benefit I did not expect and now value most. More than
once I've come to a passage sure of what it said, asked for the counter-evidence, and watched the
idea fall over. An assistant that can find the case *against* me quickly is worth far more than one
that agrees with me pleasantly.

**Depth I could not have reached alone.** Feast cycles, calendars, Ancient Near Eastern custom,
Second Temple background — tackling topics with real cultural and linguistic depth behind them has
been the best part of this. Connections I would never have found unaided are now traceable and cited.

---

## Failure modes and guardrails

| Risk | Why it matters | The guardrail |
|---|---|---|
| **Hallucination** — a fabricated lexical claim, a misquoted verse, a citation to a work that doesn't exist | This is the headline failure mode of language models, and in a Bible study it is not a small error — it is putting words in God's mouth | Research is **grounded in structured Bible databases held in this repo** — original-language text, morphology, Strong's, semantic domains, cross-references, drawn from open-licensed scholarly sources. A language claim has to resolve to a real entry in a real dataset. It is not permitted to come from recollection |
| **Fluent, confident, wrong** | Machine-generated prose *sounds* authoritative regardless of whether it is correct. Smoothness is not accuracy | Every substantive claim carries a citation, and translations are always named. An untraceable claim is a defect in the page |
| **Agreement bias** | These tools lean toward building the case you asked for. Ask leadingly and you'll get a confident yes | The study method runs exegesis *before* conclusions and requires the counter-case to be sought. "The one who states his case first seems right, until the other comes and examines him" (Proverbs 18:17, ESV) |
| **Doctrinal flattening** | A model's instincts reflect whatever is most common in its training data, which drifts toward a vague consensus and away from any defined position | Studies are written from a stated position, not a model's default. See the [Statement of Faith](https://the-way.lewy.au/about/statement-of-faith/); interpretive commitments are declared up front rather than absorbed silently |
| **Mishandling others' work** | Scholarship is someone's labour, and lifting it wholesale is theft whoever does the typing | Sources are partitioned by licence. Open-licensed data is used and quoted freely; commercial reference works are cited by name with short, attributed quotations only, never reproduced at length |

A human reads every line against the text before it is published.

---

## How a study gets made

Exegesis first — what the text meant, then and there — and only afterwards what it means here and now.

```mermaid
flowchart TD
    A[Passage or question] --> B[Historical & literary context<br>author, audience, occasion, genre]
    B --> C[Original-language word study<br>grounded in local lexical data]
    C --> D[Cross-references &<br>translation comparison]
    D --> E[Commentaries consulted last<br>to check the reading, not to form it]
    E --> F{Human review<br>every line, against Scripture}
    F -->|Doesn't hold up| B
    F -->|Verified & owned| G[Published]
```

The loop back to context matters more than the arrow forward. A conclusion that won't stand on the
passage in its own setting goes back for rework; it doesn't get published with softer wording.

---

## How it works under the hood

People ask whether this is RAG — retrieval-augmented generation, the usual pattern where you chop a
corpus into fragments, turn them into vectors, and pull back whichever ones look closest to the
question. It isn't, and the reason matters: a lot of "AI for Bible study"
projects reach for that by default and give away something they didn't need to give away.

**Scripture already has exact addressing.** Book, chapter, verse. Strong's numbers. Louw-Nida
semantic domains. I don't need a similarity score to find Matthew 24:36 — I need a key. So the
research runs on ordinary database queries against structured datasets held locally: eleven lookup
tools over Hebrew and Greek text, morphology, lemmas, semantic domains, concordance and
cross-references.

The difference is not academic. Vector search returns passages that *seem* related, ranked by
closeness, with no guarantee it found all of them. A keyed query returns the complete answer set, or
nothing. When a study here says a word occurs eight times in Matthew and four of those are in one
discourse, that is a count, not an impression — and you can re-run it. It also means an empty result
tells me something: while writing on the Olivet Discourse I wanted to claim a verbal link between
Matthew 24:35 and Revelation 21:1, and the query came back empty, because that verb never appears in
Revelation. A similarity search would have handed me something close enough to keep believing it.

**Recall is not allowed to be a source.** Anything quoted gets pulled from a database before it goes
in the file. This is not caution for its own sake — it is the failure I actually keep hitting. A
draft here once had John 6:34 as "Lord, give us this bread" where the ESV reads "Sir". More recently
I had a line of the Mishnah quoted from the wrong translation, and the wrong words came along with
the wrong edition. Both looked entirely plausible on the page.

**The method is written down, not left to the model.** The research process on this site is
Fee and Stuart's, turned into instructions the assistant has to follow — exegesis before
hermeneutics, a defined word-study method, rules about what may be quoted from where. The model
doesn't get to decide *how* a passage is studied. It gets to do the fetching inside a method I chose.

**Writing and checking are deliberately separate jobs.** There is one set of instructions for
drafting a study and a different set for auditing one, and the auditing set is told to treat every
claim as unproven until it has been re-derived from source — and not to fix anything while reading.
A reviewer who edits as he goes stops examining the parts he has already tidied. Splitting the two is
the only honest answer I have found to the obvious problem of a machine marking its own homework.

**Everything leaves a trail.** Each page records which models worked on it and when, derived from the
repository's own history rather than typed in. Behind each study there is a working file listing what
was checked, what was found, what is still open, and where a conclusion was later corrected —
including the times I overruled the assistant and the times it was right and I wasn't.

**Cheap automatic checks catch what careful reading misses.** Seventeen content checks run over every
page — frontmatter, quotation formatting, prose habits, Hebrew that would render wrong. They are
unglamorous and they find things eyes slide past.

### The part that actually does the work

None of the above is the assistant catching itself. In one week of work on this site: I found a
sentence about Christ's knowledge that was carelessly worded and would have read as denying his
deity. I knew a source was in the repository when the assistant had concluded from an out-of-date
index that it wasn't. And I suggested checking a freshly imported Bible text against an independent
website, which turned up a real bug in the import that every internal check had passed clean.

So the honest summary is not that the guardrails make the output trustworthy. **It is that a human
has to check, and the point of all this machinery is to make checking cheap enough that it actually
happens** — every claim keyed to a source you can re-query, every quotation traceable, every
correction written down. Take the human out and the whole thing stops being worth reading.

---

## Where it stops

**It is not content generation.** A study with nothing to say doesn't get written up to fill a gap in
the navigation. Volume is not the goal.

**It does not replace prayerful, thoughtful study.** AI speeds up the research; it cannot do the part
that matters. "Open my eyes, that I may behold wondrous things out of your law"
(Psalm 119:18, ESV) is not a database query. The
Spirit of truth is the one who guides into truth
(John 16:13) — a machine retrieves, He teaches.

**It has no spiritual authority.** AI is not converted, not indwelt, not accountable to anyone, and
has no discernment — only pattern. It cannot pray, cannot be convicted of sin, and cannot be taught
by God. Nothing it produces carries weight because it produced it.

**The wrestling is not outsourced.** Sitting with a hard passage until it changes *you* is a large
part of why anyone should study Scripture at all. Hand that over to a tool and you end up with a
better page and a poorer soul. The tool does the fetching, the collating, and the checking; the
sitting with it stays mine, and so does the working out of what it demands.

**Errors here are mine.** If something on this site is wrong, I published it.

---

## Checking the work

The Bereans were commended for checking the Apostle Paul himself against the text
(Acts 17:11 (ESV) — "examining the Scriptures daily
to see if these things were so"). If Paul's teaching was fair game for verification, so is this.

**Test everything; hold fast what is good**
(1 Thessalonians 5:21, ESV). The citations on these
studies are there to be followed. If something is wrong,
[tell me](https://github.com/ding0t/bible_studies/issues).

Scripture is the authority. This site is one person's working out of it, assisted by a tool that is
very good at retrieval and utterly incapable of faith.

---

## Resources

- [AI’s Usefulness and Its Dangers for Preachers, The Gospel Coalition, 2025-10-14](https://www.thegospelcoalition.org/article/ai-usefulness-dangers-preachers/)
    - 2 Timothy 2:14