AI in Bible Translation?

No AI was hurt in the process of writing this article.

It is indisputable that technology has been helpful for humans, though it comes with trade-offs. It seems to me that usually the first things to get automated are the more mechanical or non-creative parts of a task. The creative work is left for humans. AI (as in LLMs) crosses an important threshold of showing creativity (emergent abilities), mostly by mimicking human language. AI is, therefore, now playing a bigger role in the creative aspects of human work, like prose and song composition.

Translation as a task is both mechanical and creative. It is mechanical to ensure that meaning is preserved across languages. It is creative in that the translator has a choice in how they express that meaning appropriately for the target community. Now, it is not unusual to use a spell checker or a punctuation checker for Bible translation since their contributions seem benign. But as suggested earlier, LLMs (and generally machine translation suggestions/drafts) encroach upon the creative territory that has been conventionally earmarked for humans and it is, therefore, valuable to investigate what this means for Bible translation.

We could draw a parallel to what is happening in the world of programming languages today. It was not until a few years back that almost all code was hand-written. I say "almost" because there were tools that automatically generated boilerplate code or quickly bootstrapped a software project from a predefined template. Moreover, Integrated Development Environments (IDEs) have some fancy tooling for code suggestions and quick-fixes. None of these, however, came close to what AI (especially agentic) coding models can do today. Claude could code up an app from scratch with just a single prompt. Or it could even enable your dog to develop games (no more 'my dog ate my homework' but 'it actually did my homework!'). Whether these are comparable to a production-quality application is a separate discussion. The very speed at which you can develop these makes them a separate category to consider seriously.

Anything I say about this has the risk of aging poorly. But at the moment, AI models are becoming an increasingly prominent part of software development in the industry. The best practices for using AI in computer programming are still in their formative stages. Not only because it is a rather recent phenomenon, but the AI models themselves seem to be doing better/more every few months. One common refrain I have heard is to treat the AI model like an intern (not used pejoratively; I was an intern once). That is to say that the human should clearly document requirements (inputs, outputs, restrictions, etc.), participate in breaking down the project into tractable and measurable steps, and provide a reliable testing harness for the AI model to self-evaluate its progress against. This also means that at the time of writing, we do not generally expect AI models to code a larger body of work to a human's satisfaction without consistent steering or intervention. Software developers have many opinions about how code should be written and structured for a system that is in production (i.e. a system gainfully used by more than a few people). This is because the developer takes responsibility for the artifacts they produce (code, documentation, etc.) so that the system actually works. This means they treat the AI models ultimately as a tool and are motivated to scrutinize its output since they are the ones ultimately signing off on what is produced and run.

There are similarities in writing code and translating the Bible¹: both care about what is written/expressed; there are various checks done and changes are carefully considered. Similarly, I believe it would benefit us in the Bible translation space to also view AI models as a tool that needs the translator to ultimately steer and guide it. More importantly, the translator should take responsibility for any AI contributions. This changes the question from 'Am I allowed to use AI?' to 'Is the AI output useful?'—which allows us to move forward productively without ceding control. I understand there are concerns about AI drafts causing an anchoring effect. Though reasonable, some of the issues are mitigated by having verses go through multiple rounds of checking with different translators and ultimately having the community vet the translation. Another risk is not providing enough context/training to translators on the field who are not tech-savvy and may never have seen AI models previously. The danger is that an unsuspecting translator may end up trusting the tool too much and forego their own valuable judgment in the work. This is a danger that anyone using AI today has to wrestle with. The tool developers need to take this into account and build guardrails to protect the translators. This could be through features that require verification of AI drafts or asking comprehension questions. I believe Bible translation is inherently spiritual and God has commissioned the Church (the body of Christ) to do this work. As long as we, as believers, are taking responsibility and ownership, using a tool (as powerful as it may be) keeps us accountable still to Christ's calling on our lives. Shifting the focus to translation quality and having translators take ownership of the output allows us to focus on what finally matters—having the Word of God accessible to every language group² (Rev. 7:9).

Footnotes

¹ I would be remiss to not state that there are significant ways in which computer programming is different from Bible translation. A computer program is ultimately defined by its behavior, i.e., the inputs and corresponding outputs. From this perspective, we could swap out the code for something else (e.g. different programming language) as long as the input-to-output contract is maintained. This is the basis for how LLMs can now write code that is increasingly treated as a blackbox to the point that there are enthusiasts who don't care what the code is anymore as long as the desired output/outcomes are achieved. We cannot, however, afford this for Bible translation—the Word of God must be accurately translated, and we do not have the freedom to add or subtract from it.

² The ETEN Innovation Lab is committed to meeting the All Access Goals, which is in alignment with the Biblical mandate.