Theological Bias in AI-Generated and AI-Checked Translations: Ethical Implications

This paper investigates the effects of bias in large language models (LLMs). The issue of bias in LLMs is a well-documented phenomenon; it results from prejudice, partiality, or particular views latent in training datasets. This contribution explores ethical concerns related to the appropriate use of these models for the tasks of drafting, checking, and other automatable steps of the Bible translation process. In particular, it examines (1) differences in machine translation choices and theological assessments made by LLMs trained on datasets with known biases, like those compiled by Jehovah’s Witnesses; and (2) differences in these same choices by models fine-tuned toward particular ideological, philosophical or theological positions. Understanding these differences leads to better design and safety measures for AI translator copilot tools, as explored by Whitenack et al. (2023), and further informs concerns about the loss of “sanctity between the translators and the Holy Spirit” during the drafting process, as raised by de Blois et al. (2023). Finally, the paper inquires into potential bias mitigation strategies when using LLMs for automated drafting and theological checking, such as the appropriate transparency of training datasets and the evaluation of these datasets for types and levels of latent bias.