Generating Wisdom: AI and the Bible Conference Abstracts

How do we think as Christians about the opportunities and challenges of using AI in the work of Bible translation, biblical exegesis, and biblical hermeneutics? And how does the biblical witness itself provide resources for answering this question? Rather than argue for what people of the book “ought” to do with AI, this presentation explores the possibilities and pitfalls of AI-assisted translation by locating it within the larger story of the biblical witness. In doing so, it identifies a few key plot points in the biblical story that might serve as theological signposts for anyone using or exploring AI as a technology and/or a tool, and even more specifically, on large language models that engage in natural language processing tasks.

Our quest to build intelligent machines can produce valuable insights for theological anthropology through both its successes and its notorious failures. This paper outlines how a zoomed-out theological reflection on the history of AI research so far can foster a necessary breakthrough in the theological interpretation of the biblical notion of imago Dei. Evolutionary theory spurred a complexification of the theological discourse around what it means to be human and in the image of God, pushing the interpretation of these notions away from a substantive understanding and toward functional, relational, and eschatological formulations. Similarly, I argue, engagement with AI can provide new “data” for theological theory, triggering a refinement of our interpretation of human distinctiveness and imago Dei. Contemplating what might still render humans distinctive in a hypothetical scenario of human-level AI leads to an appreciation of the phenomenological dimension of human nature, and it points to our unique relationship with God as the most fundamental aspect of the divine image. This has profound implications for our assessment of claims made by chatbots powered by large language models, and the paper reviews some of these implications.

A pressing reason for the development of artificial intelligence in our Information Age is to improve information flow. For the book of James, according to Joel Green, our ways of embodying information are formed according to a Christological narrative that runs from creation to new creation. Green’s work on James and the methods of theological interpretation I’ll explicate concern the ethics of information flow. Patterns of information flow should be conservative of pre-existing norms, contextualized rather than abstracted or generalized, and communicative rather than transactional. These patterns, as some information ethicists agree, allow for robust innovation of artificial intelligence even while maintaining the systems of personal and social formation upon which we depend.

This paper offers a biblical response to artificial intelligence by focusing on generative AI and, more specifically, the role of large language models (LLMs). With a history covering many millennia, peoples, and a persistent God, the Bible is adept at identifying issues around the misuse and confusion of language, as well as its foundational and creative nature. Given the ongoing technical issues with LLMs, such as bias, hallucination, and the alignment problem, this paper suggests that biblical wisdom regarding language may be opportune. Examining two familiar biblical narratives, the tower of Babel and the post-resurrection Pentecost, the paper first addresses the power of language in bringing together communities while also dealing with the challenges posed by the diversity of languages and peoples. How might this inform those creating policies around the diversity of training data and finetuning LLMs? Second, the paper explores the context and canonical history of these biblical texts. The structure of the Bible is often taken for granted, but it was a process that involved arduous debates and spirit-filled moments that brought texts into canonical alignment, with some books chosen and others left out. How do we wisely select data and provide an evolving system to address what often looks unbounded, confused, and in need of care?

In some sense, ancient apocalypses were a means of generative creation; coupled with imagination, hearers and readers understood the future of the world. In this talk, we will examine the meanings of and basis for ancient apocalypse, affirm the value of theological imagination, and consider how this feeds into our perception of the future of AI. Observing and then disavowing ourselves of the common assumptions of AI as a harbinger of the apocalypse that John describes in Revelation, we will turn instead to the uniqueness of AI and what it may say about ourselves, God, and the new heaven and the new earth. We will conclude with hopeful possibilities that AI may lead to the fulfillment of apocalypse (in a good way).

Artificial intelligence (AI) and other computational and statistical methods enable new ways to interpret religious texts. In these methods, the meanings of words generally depend upon their network of associations with other words. Some approaches, like the large language models (LLM) underlying recent advances in AI, then ground these meanings in large samples of text. Other methods use only the text itself, like topic modeling, which uses the word colocations to identify topics, or themes, within the text. The former, distributive approaches can bring much online knowledge to bear in text interpretation, though that includes biases, unreflective assumptions, and a tendency toward the generic. The later, associative methods can provide a fresh perspective on texts and can augment human interpretation.

One of generative AI's great strengths is the ability not just to query databases, but to speak to them using natural language. Jewish sacred texts, which pious coders have been digitizing since almost the dawn of computing, are therefore only a small step away from being in direct conversation with the people who wish to study them. This possibility promises to revolutionize the study of Jewish texts, dramatically lowering the bar to entry. At the same time, the idea of conversing directly with the Torah—or a biblical figure, or a medieval commentator—carries with it the risk of distortion, and of changing the locus of authoritative messaging. Understanding that this transformation is probably inevitable, what will happen when AI-driven manifestations of scripture become widely available—and how should they be regulated?

Jesus laid upon the early church the Great Commission, a mandate to go make disciples of all nations. We are to use all means at our disposal to carry this out, which today includes the tools of artificial intelligence. Within the world of Bible translation, there is a united push to ensure nearly 100% of the world access to the Bible in their heart language by 2033. In order to do so, we need to harness the power of AI to accelerate the pace of Bible translation. This talk lays out boundaries for what we should ask of AI (because it can do it most efficiently), what we should not currently ask of AI (because it is not yet capable of doing it well), and what we must never ask of AI (because it dehumanizes the process of Bible translation). Questions of contextualization, local theologizing, and community ownership of a translation will all be addressed within a framework of enthusiastically embracing AI for all it can do well.

Efficiently and accurately translating a corpus into a low-resource language remains a challenge, regardless of the strategies employed. Many Christian organizations are dedicated to the task of translating the Holy Bible into languages that lack a modern translation. Bible translation (BT) work is currently underway for over 3,000 extremely low resource languages. We introduce the eBible corpus: a dataset containing 1,009 open-source partial or full Bible translations spanning 833 different languages and 75 language families.

In addition to a BT benchmark dataset, we introduce 168 model performance benchmarks built on the No Language Left Behind (NLLB) neural machine translation (NMT) models. Finally, we describe several problems specific to the domain of BT and consider how the established data and model benchmarks might be used for future translation efforts. For a BT task trained with NLLB, Austronesian and Trans-New Guinea language families achieve 35.1 and 31.6 BLEU scores respectively, which spurs future innovations for NMT for low-resource languages in Papua New Guinea.

What concrete questions should we ask when applying AI to help solve specific Bible translation problems? The Bible translation world is collaborating on Project Accelerate, an initiative to build tools powered by emerging AI technology and established natural language processing (NLP) techniques. Its vision is to make intuitive translation software that recedes into the background and keeps the translation project the center of focus. It is targeted to address two significant issues with existing Bible translation tools. First, user interfaces are too complex, making it hard for translators to use existing tools. Second, software tools are isolated, leading to duplication of effort. In this presentation, we will probe how this project is working to answer the right questions at the right time to apply AI to accelerate Bible translation wisely. It will serve as a case study into the kind of questions and challenges that an ongoing project should answer in its quest to apply AI successfully to advance Bible translation, interpretation, and beyond.

We are seeing a recent explosion of interest in AI-driven tools that are paired as “copilots” with human workers. These include copilots that are helping users to write code, find sales leads, draft legal documents, generate design ideas, create lesson plans, and much more. Could such tools also be of use in the translation of scripture? This talk introduces AQuA (Augmented Quality Assessment), an AI copilot developed by SIL International to help stakeholders in the Bible translation process more thoroughly and efficiently evaluate certain qualities of translation drafts within the time available. 

AI tools, particularly large language models (LLMs), offer promising prospects for Bible translation, but also present unique challenges. Machine translation tools, while powerful, rely on imitation rather than creativity and often function as a "black box," making the process opaque. The social and contextual nature of human translation is often overlooked in machine translation. By recognizing these limitations and designing translation software to mimic social processes, AI can be used as an assistant rather than a replacement, enhancing the quality of translations. The future of AI in Bible translation should prioritize open, socially informed processes.

New digital technologies provide expanded access to Jewish wisdom by enabling us to build pathways into canonical Jewish texts for those interested in exploring them. Machine learning and artificial intelligence allow curious laypeople, even those without backgrounds in Jewish text study, to locate, read, and understand books that were previously difficult or impossible to approach. Strategies such as generating links between texts, auto-translation, topical tagging, scaffolding materials written by LLMs, and retrieval-augmented generation are being deployed to open up the texts of the Jewish tradition to those who were previously shut out. This session will explore some of the ways these techniques are being used and their implications for opening up ancient Jewish texts to modern readers. 

With these new technologies come new concerns about the goals and perils of increased access. Deploying these tools requires us to consider a broad range of questions, from how to ensure quality control and how and when to be transparent about sources of information, as well as the deeper ethical challenges: What happens when access to information is democratized? How might wider consumption of these texts change the balance of power among scholars and laypeople, and what is our responsibility to address those changes? What are the costs of this greatly expanded access, and how might we assess when it is worthwhile?