Filling the gap in manuscript transcriptions
The transcriptions I have stop short of some periods the study needs. Through the end of the year I am working with other manuscript researchers to fill that gap.
The research
AI and New Testament textual criticism
My PhD research asks whether AI-assisted analysis can measure intrinsic probability in the 233 hardest textual decisions of the UBS6 Greek New Testament. Here is the question, the approach and where it stands, written for curious readers, with a note for specialists.
The question
The editors of the UBS Greek New Testament grade each of their decisions from {A}, certain, to {D}, very difficult. In the 233 decisions they rated {C}, they had difficulty deciding which reading to print.
One factor in every such decision is intrinsic probability: which reading best fits the author's own vocabulary, style and thought. That factor has always rested on a scholar's judgment. I'm testing whether AI-assisted computational analysis can ground it in measurable evidence, so it can carry more weight than it does today.
Why it matters
The Greek text behind modern English Bibles rests on thousands of these decisions. Methods that can be measured and repeated are easier to check, to teach, and to trust.
Behind the work is a conviction that how God has preserved his Word is worth studying carefully.
τὸ δὲ ῥῆμα κυρίου μένει εἰς τὸν αἰῶνα · 1 Pet 1:25
How I'm going about it
Gather the hard cases
Start from all 233 variation units the UBS6 editors rated {C}, with the manuscript evidence for each.
Measure each reading
Score how well each competing wording fits its author, with several instruments: language models I train myself and others, all run locally.
Keep what holds up
Apply many methods, across all the units and in depth on a few, and report only those that prove empirical and repeatable.
For specialists: the study works within reasoned eclecticism, from UBS6 and Houghton's Textual Commentary (2025). Its scope is intrinsic probability only; transcriptional and external evidence are outside it.
Where things stand
- Done Began the PhD in Biblical Studies, Liberty University Fall 2025
- Done Research plan presented Oct 2026
- Done Thesis and scope set Oct 2026
- Done Witness explorer built Oct 2026
- Now Filling gaps in the manuscript transcriptions Through Dec 2026
- Now Drafting Chapter 2, the history of the method Fall 2026
- Next Full dissertation proposal Due Nov 2026
- Next Build the measurement codebook and a pilot To come
- Next Train and run the models across all 233 units To come
- Next Write up the results and defend To come
Common questions
What is New Testament textual criticism?
The New Testament survives in thousands of handwritten Greek manuscripts, and no two agree in every word. Textual criticism compares them to judge, at each place where they differ, which wording the author most likely wrote. Modern editions of the Greek New Testament, and the English Bibles translated from them, rest on those judgments.
What is intrinsic probability?
Editors ask two kinds of question about each competing reading. Transcriptional probability asks what a copyist was likely to have changed. Intrinsic probability asks what the author was likely to have written, judged from his vocabulary, style and thought elsewhere. This study is about the second.
What does a {C} rating mean in the UBS Greek New Testament?
The United Bible Societies' Greek New Testament grades each textual decision it reports from {A} to {D}. A {C} means the committee had difficulty deciding which reading to place in the text. UBS6 has 233 of them, and this study takes up every one.
How does the study use AI?
Language models learn from each author's own Greek writings, then score how well each competing reading fits that author. Several kinds of model are used, including models I train myself, all run on my own computer. Many methods are compared, and only those that prove empirical and repeatable are reported.
Will a computer decide the text of the New Testament?
No. Intrinsic probability is one factor among several in an editor's decision, alongside the manuscript evidence and what copyists were likely to change. The study asks whether that one factor can be measured well enough to deserve more weight than it gets today.
All updates
A witness explorer for the study
I built a tool that reads the UBS6 text beside its manuscript evidence and pulls out each witness and each author's writings as clean text for the models.
Many methods, several kinds of models
Rather than trust one approach, the study applies many methods and reports which prove repeatable, testing models I train myself alongside others.
The thesis is set
The dissertation will argue that AI-assisted analysis can make intrinsic probability measurable enough to merit more weight than it receives today, tested on all 233 {C}-rated variation units in UBS6.
Research plan presented
I presented the plan in a doctoral research seminar. The feedback widened the scope from a sample of units to all 233.
Get new updates by email
A short note each time I post. No gift needed; unsubscribe anytime.
Ask me a question
About the method, the manuscripts, or the work in general. I read every message and reply as soon as I can.
Your email is used only to reply to you. See the privacy note.
Want to help this work move faster?
Secure checkout by Stripe, through my company, JARVIS Consulting, LLC. Gifts are not tax-deductible.