The Use of Artificial Intelligence in Bible Translation

From Justin Goldberg - Linguistic Advisor with SIL Ethiopia (11 August 2026)

This document (below) is a summary of presentations and discussions about the use of Artificial Intelligence (AI) in Bible translation that took place at the Bible Translation Conference in Chiang Mai, October 28-31, 2025. A list of papers presented at the conference related to AI can be found at the end of this document. I have summarized the papers presented at this conference under four headings: 1) what AI is and is not, 2) the ethical use of AI, 3) AI Tools currently available, and 4) views on the use of AI for drafting Scripture.

---

Extracted text from The_Use_of_Artificial_Intelligence_in_Bible_Translation.pdf:

The Use of Artificial Intelligence in Bible Translation A Summary from the 2025 Bible Translation Conference

Justin Goldberg

Linguistic Advisor SIL Ethiopia

August 11, 2026

This document is a summary of presentations and discussions about the use of Artificial Intelligence (AI) in Bible translation that took place at the Bible Translation Conference in Chiang Mai, October 28-31, 2025. A list of papers presented at the conference related to AI can be found at the end of this document. AI use in Bible translation is clearly a prominent topic on the minds of many involved in the Bible Translation Movement. This was especially clear at the most recent Bible Translation Conference, given the large number of papers presented on the topic. There was a wide range of views presented, from practical and technical to philosophical and theological. It should be noted that while there are many projects around the world using AI, not everyone agrees on the extent to which it should be used. While some are more driven by the practical advantages of what is technically possible, others are expressing caution over the possible harms that can easily be overlooked. Most people hold some combination of both of these views. In any case, it was clearly stated multiple times at this conference that AI use in translation is in the experimental and developmental stages. I have summarized the papers presented at this conference under four headings: 1) what AI is and is not, 2) the ethical use of AI, 3) AI Tools currently available, and 4) views on the use of AI for drafting Scripture.

1. What AI is and is not Before summarizing conference presentations on ethics, practical uses, or differing views on AI, it is important to have a common understanding for what AI is and is not. Many people

presenting papers spent time to make this clear to avoid unnecessary arguments and misunderstandings. AI is a tool; a tool for doing exegesis and researching biblical context, for checking drafts for errors and inconsistencies, and for providing a form of a draft. As a tool, it can be used skillfully or unskillfully by the human user of the tool. In plain terms, AI is an advanced computer program that can find and replicate patterns in language. This capability can be harnessed for different purposes, including quickly scanning biblical resources and providing answers to exegetical questions, comparing a draft to another translation and highlighting places where there are differences, or producing a text that is based on high probabilities of words based on established patterns (AI drafting). AI by definition (artificial) is not human. As a tool, it cannot be inspired by the Holy Spirit (that we are aware of). It is not intended to replace human translators. It is not able to understand meaning; it only recognizes patterns. It is not ‘out to get us’.

2. Ethical use of AI1 Since the theme of this year’s conference was ‘Ethics in Bible Translation,’ there was a lot of information presented on ethics. I heard from a number of individuals that this conference was much more robust in dealing with both the positives and negatives of AI use. These individuals said that the 2023 conference had many presentations on AI drafting, but was mostly focused on the exciting possibilities without considering the effects on community acceptability and on the translators themselves. Those two ethical areas are summarized below under the headings Stakeholder Approval and Mitigating Harm. 2.1 Stakeholder Approval One of the major points on using AI ethically in a translation project was that stakeholders need to understand the positive and negative possible outcomes of using AI in the translation project. Much of the time on this topic was spent focusing on the language community and making them aware of these outcomes. It is not enough for a few reviewers or community connections to give their approval, because community acceptability and ownership is 1

SIL has produced an ethics statement, available at https://ai.sil.org/ethics.

fundamental to the translation being used effectively in the community. This poses a problem for translations for remote people groups with limited exposure to technology. How can they really understand AI, when very few people with a lifetime of technology use understand it well? While the difficulty in communicating these things well was admitted during the conference, it was also stressed that attempts need to be made to help the community understand what possible outcomes could be. Below is a chart that provides some of the positive and negative possible outcomes for using AI in translation that were discussed at the conference. This is not meant to be an exhaustive list. Positive

Negative

Speed: drafts can be produced much faster,

Mistakes: There will always be mistakes

in fewer years. AI also streamlines

consider this less acceptable from a machine

resulting in finishing the translation project workflows so that tedious tasks do not take up so much time, and more time can be

produced by AI drafts. The community may than from a human. Tests have also shown

that translators editing AI drafts have lower

allocated to things like exegesis and checking comprehension of the text than those who sessions.

don’t use AI drafts.

Roles retained: translators are not being

Roles changed: translators are now editors,

replaced by AI; they keep their positions and

and if they are not very gifted at detailed

translation.

missed.

their role is still vital to producing a good

copy-editing, the AI mistakes will likely be

Consistency: AI drafts will always use

Translator development: due to lower

terms tool in Paratext. This means that AI

negative impact on spiritual formation. The

established biblical terms from the biblical drafts are more likely to be consistent in rendering these terms in the draft.

2

comprehension of the text, AI can have a community may want to prioritize the

development of their translators or make sure that steps are being taken to ensure that this gap is addressed.

2

However, if translation teams have multiple entries for one biblical term, then the AI draft might be less consistent. It is common for teams to input multiple terms for one biblical term since most words have multiple senses (different meanings depending on context). If there is enough data, then the AI should not have a problem using the correct term based on contextual clues.

Access to knowledge: AI provides access to

Hallucinations: when the AI model is not

accessible.

fills in that gap with a ‘guess’; people

resources that would not otherwise be

able to find an equivalent for a translation, it working on the AI draft currently have no way of knowing what might be a hallucination.

2.2 Mitigating Harm With all the amazing potential that AI offers, there are clearly negative outcomes, as can be seen in the chart above. It was encouraging to see at this conference that developers of AI software were seriously evaluating the impact that AI use has on translators. Without awareness of these negative impacts, nothing will be done to mitigate harm. With awareness comes the ability to take steps to mitigate harm. Being aware of the harm and not taking steps to mitigate it is unethical. For example, since we know that translators comprehend less about the text when using AI drafts, steps can be taken by project leadership to help strengthen this area.

3. AI tools currently available There are many AI tools already available for different kinds of tasks. I have arranged them into three groups: exegesis (3.1), drafting (3.2), and checking (3.3). This list will undoubtedly grow rapidly in the coming years. A good way to stay up-to-date on these changes is to check the tools.bible3 website, which was designed to be a centralized library for all translation tools.4 Large language models (LLM’s) such as ChatGPT and Claude are excluded from this list; they can be useful tools for exegesis and checking but because they draw from all possible internet sources without discernment, the responses may be less trustworthy than those listed below. 3.1 For Exegesis

3

tools.bible For the purposes of this document, I have limited the list of AI tools to those that were discussed in presentations at the BT Conference. Looking at the tools.bible library, it is clear that there are other tools that have been developed not discussed here, such as translationCore, Lilt, and Codex Translation Editor. 4

FaithBridge5: this tool is an AI chatbot that allows users to type any question into a text box. In the Bible Translation Chat tab, the chatbot references databases of curated biblical resources 6 and provides answers to the questions. Since these databases are curated, inappropriate information is not referenced in providing answers (such as biblical resources from a Jehovah’s Witness perspective that could be included when using LLM’s like ChatGPT). Follow-up questions or other prompts can be added to fine-tune the results, and suggestions for how to translate materials can also be provided. There is also an Images and Videos tab that allows users to find media to help them understand or visualize things they are translating. Finally, there is a Dynamic Retelling tab that allows users to read and hear a passage of Scripture translated in a dynamic way to help understand the meaning better. 3.2 For Drafting Scripture Forge7: by far the BT Conference’s most discussed platform for AI drafting, Scripture Forge (SF) provides web-based and app-based translation services. Teams can choose to only use the SF platform and do all drafting and checking on that platform. For teams using Paratext, SF syncs with Paratext, using already drafted portions as the source text when producing new AI drafts. After SF creates an AI draft, it can be synced with Paratext so that the draft appears in the project text window or as a Scripture resource. Since the SF draft is trained using already checked project drafts, it is important to already have a number of books completed using traditional methods that have been well-checked and approved by the team as being of high quality. This often means that teams already have a completed NT and they plan to use SF for AI drafting of OT books. Avodah8: Avodah is different than Scripture Forge in that the AI drafting model is trained by the translators over a period of up to two years. The team begins by translating a selection of 1200 verses designed to introduce the model to the language. Texts and grammar materials can also be added to assist the model. Then the model produces the first drafts of Scripture. These

5

faithbridge.multilingualai.com These databases include Bible Aquifer (aquifer.bible), the Translation Insights and Perspectives website (https://tips.translation.bible), and the MACULA Greek (https://github.com/Clear-Bible/macula-greek) and Hebrew (https://github.com/Clear-Bible/macula-hebrew) datasets. 7 scriptureforge.org 8 https://www.avodahconnect.com/bible-translation 6

drafts are very rough to begin with, so the translators identify each error and re-train the model. Then another draft is produced that should be improved from the previous draft and the process is repeated until the drafts being produced are accepted as being a good starting point for the team to work with. Then the team moves forward with the translation process, doing their normal checking stages for each book after having this initial AI draft. The model is designed to ask questions to verify things that it has learned. Aero9: designed specifically for oral Bible translation (OBT) projects, Aero is a collection of AI tools that addresses audio-specific challenges. It can remove unwanted background noise, change the quality of a person’s voice, replace a key term everywhere it occurs in the recording, transcribe audio into a written form, and convert text to speech in the project language. Audio AI drafting capabilities are currently in development. 3.3 For Checking AQuA10: AquA (short for ‘Augmented Quality Assessment’) is designed to augment the project’s quality assurance processes (checking for accuracy and naturalness) by comparing the project text to another translation. After comparing the two texts, AQuA provides visualizations of five different categories: word correspondence, semantic similarity, text lengths, readability, and repeated phrases within in the project and comparison text. These visualizations allow the team to see where they are translating more or less word-for-word (word correspondence), more or less dynamic (semantic similarity), more or less wordy (text lengths), longer or shorter sentences (readability), and whether or not repeated parallel texts match each other (repeated phrases). Users can also search for terms to check for consistency between the two texts and for omissions. In the future, AQuA will provide comprehension questions to assist with ensuring clarity of the text.

4

Views on the use of AI for drafting

Everyone at the Conference had the good intention of improving the translation process and seeing lives transformed. However, not everyone agreed on the best way to use AI for drafting

9

https://ai.sil.org/projects/AERO https://aqua.multilingualai.com/

10

Scripture. Four views presented are summarized below: AI draft as first draft (4.1), AI draft as draft zero (4.2), AI draft as a reference Bible (4.3), and AI draft as a drafting check (4.4). 4.1 AI draft as first draft With AI models such as the one used by Scripture Forge, the AI draft appears in the translation software as if one of the translators had drafted it themselves. The translation team begins Team Check using this AI draft. No one at the conference suggested this was the best option; but it was presented as an option for teams to consider. 4.2 AI draft as draft zero As with the AI draft as first draft, the draft appears in the translation software as the main draft. However, it is not considered to be a first draft yet. First, team members must evaluate the draft, making edits that will help to influence better accuracy in future AI drafts. At the end of this editing stage, it is considered a first draft, which can then go through Team Check. The only difference between this view and the previous view is that there are two checking stages after the AI produces a draft (editing and Team Check) whereas there is only one checking stage in the previous view (Team Check). Therefore, these two views are essentially the same. 4.3 AI draft as a reference Bible The AI draft is used as a reference Bible along with traditional translations from languages of wider communication (LWC’s) like English and Amharic. The draft window in the translation software remains blank and the translator builds the draft themselves. The AI draft is given a unique project name in the translation software and added to the text collection window along with other LWC Bibles as a reference to help understand meaning. More than one AI draft can be created based on different source translations so that the translators have access to more or less dynamic versions in their language. The main idea in this view is that the human translator remains the one making the first draft. However, they are still influenced by the AI draft, often before they have actually drafted anything on their own. 4.4 AI draft as a drafting check (or AI completely rejected for drafting) The AI draft is used as a reference as in the previous view, except that it can only be referenced after the translators have produced a first draft. This view rejects using AI for a first

draft, draft zero, or drafting help for both technical and spiritual reasons. Technically, current AI drafting software is not a translation of meaning, but a sentence-by-sentence arrangement of patterns. This means that statistically higher-use terms and phrases will be preferred over others, without the ability to use the context as a guide. Also, AI drafting will make guesses when it doesn’t have enough data to find a pattern. Translators have no idea whether a word or phrase in a draft is based on an established pattern or a guess. Spiritually, AI is a tool created by humans, and therefore not similar to a human guided by the Holy Spirit. It can make things look like good, natural language, but it cannot wrestle with the meaning of Scripture. This means translations produced using AI drafting cannot be considered ‘Meaning-based’.11

5

Final thoughts

One other thing that was stressed by many presenters was that there needs to be considerable effort put into assessing readiness for using AI. Stakeholders of all types should be part of this effort. Some of the most important criteria for assessing readiness are as follows: 1) the community being served by the translation needs to be aware of the pros and cons of using AI and approve of its use, 2) large portions of Scripture should already have been drafted using traditional methods, 3) the translation team should understand and accept how their roles will change, and 4) translators editing AI drafts need to have a very high level of skill in critical thinking and editing. If all of these criteria are met, there should still be further discussion about whether or not AI will be used for drafting, and how to use it well in other ways to improve the translation process. Since the use of AI for translating the Bible is still in early stages, care should be taken to regularly evaluate whether or not it is serving the project well and how its use should be adjusted. This requires flexibility on the part of all stakeholders; the promise of AI as a tool to improve translation speed and quality is undeniable, but there are very real potential dangers that need to be understood and addressed. Failure to do so has the potential to undermine not only the quality and accuracy of a translation, but also its acceptability. A Bible that is not accepted by the community cannot have an impact, whether it was completed in one year or thirty.

11

Additional resources that were used to support this point: Bender et al. 2021, Mollick 2024, Autor and Manyaka 2025, Crouch 2024)

6

Papers presented

Below are the papers presented at the 2025 Bible Translation Conference in Chiang Mai, Thailand. Many of the papers and/or recorded presentations are available at the Conference website.12 Tuesday, October 28

A Case Study of FLExTrans Performance on the Field (Ron Lockwood) A Study of Equivalence in Methods of Using AI-Generated Drafts in Bible Translation (David Duncan, et al)

AI Drafting of Scriptures: A Detailed Analysis of the Results in a Language in Papua New Guinea (Martha Wade) Community Intelligence Meets AI: Can Local Literacy Programs Improve AI-Generated Scripture?

(Taeho Jang, et al)

Finding a Place of Computer Generated Drafts in the Scripture Drafting Process (Roger Stone, David Duncan)

Integration of Scripture Forge into Bible Translation Projects (Jeff Shrum) Ethical Pitfalls of AI Drafting: Insights from Cognitive Linguistics & Philosophy of Mind (Ryan Douglas Kopke)

Why Bible Translation Organizations Need Digital Literacy Training with Biblical Ethics (Sheryl Silzer) Wednesday, October 29

Steering From Harm: Applying an Ethical Research Lens to AI and Bible Translation (Eshinee Veith)

More Than Machines: Human-Centered Ethics in AI in Bible Translation (Joshua Nemecek, Cassie Weishaupt)

AI and Mother-Tongue Translators: Navigating Ethical and Practical Challenges in Bible Translation (Margaret Muthwii)

12

https://btconference.org/proceedings

Key Term Research in the AI Era: Relevance and Impact (Minie Mathew) Ready or Not, Here AI Comes! Practical Considerations for Ethical AI Drafting in Bible Translation

(Taeho Jang, et al)

A Framework for Understanding and Implementing AI Ethically in Bible Translation (Joe Burgin) Thursday, October 30

Is the Holy Ghost in the Machine? Ethical Considerations for Using AI in Bible Translation (Damien Daspit, et al)

More Than Consumers: Parsing the Claims of AI Through the ‘Thick’ Lens of Our Ancient Theology (Myles Leitch) Theological Bias in AI-Generated and AI-Checked Translations: Ethical Implications (Ian Lindsley)

AI and Emerging Technologies in Bible Translation (Discussion Group) Friday, October 31

The Privilege of First Draft: Technical and Ethical Aspects of Using AI to Draft Scripture (Judy Heath, et al)