www.thegospelcoalition.org
My Manifesto on AI and Religion
I’ve been watching, taking notes, writing, and talking privately about the intersection of AI and religion for almost four years. Over that same period, I have participated in extensive discussions with more than 200 AI engineers, tech leaders, policy makers, and religious leaders. I have accumulated hundreds of thousands of words in my head about the best approach to religious questions for AI. I’ve condensed those thoughts down to brass tacks in this mercifully short manifesto on AI and religion written for the leaders behind those foundational AI models.
Right now, there’s one especially significant problem in AI and religion: How do we get reliable and thorough answers for every religious tradition from the AI models?
So far, the results have been mixed on the reliability and thoroughness of answers. But this problem is solvable.
My thesis is simple: Answers to religious prompts—especially those with little to no prompt engineering—can be rapidly improved by (1) better training directly from the world religions themselves, and (2) applying a principled pluralist framework for religious freedom in the foundational AI models.
Epistemology and Religious Trust
Humanity is facing a significant problem that we’ve not encountered to this point. We’re about to go from a species that acquired knowledge through primary sources to a species that acquires knowledge through secondary sources. This moment places foundational AI models in a kind of high priestly epistemological seat. This is both good news and bad news for the foundational AI models.
Typically, AI corporations think their biggest choke points in the tech stack are silicon, memory, water, and energy. But there is a fifth choke point—public sentiment. On the one hand, there’s tremendous opportunity for progress and profit. But on the other hand, there’s tremendous risk of destroying the sentiment and public trust capital needed to grow and sustain that progress and profit.
In one of the largest religion surveys ever conducted, Ryan Burge and I discovered that, perhaps surprisingly, people who are actively religious are the ones who have the highest confidence in AI right now:
Imagine the year is 2028 and a major world religion files a religious bias class action lawsuit on behalf of all adherents of that religion and names many of the foundational AI models as defendants. The suit claims that the models biased one religion over another and/or made inaccurate or incomplete statements about the fundamental tenets of that religion. That case would force all sorts of interesting legal arguments surrounding the First Amendment and Section 230. Nevertheless, this scenario is probably something these corporations would want to avoid. Currently I believe many of the models are at risk.
Simple Solution
As humanity goes from acquiring knowledge from primary sources to secondary sources via generative AI, it’s critical that religious knowledge be transferred accurately and thoroughly for any and all traditions. The free exchange of accurate and thorough religious ideas is critical to both freedom of speech and religious freedom. On these grounds I argue for a principled pluralist framework that allows every religious tradition to accurately and thoroughly share its core tenets.
Here’s how this can be accomplished. First, foundational models can work directly with top scholars of each major world religion to receive better training sets, especially as it pertains to essential, non-contested, and universally accepted aspects of each major religious tradition or group. While one might observe that this training data is already in the models, there is still confusion at points on key technical terms. One of the many challenges are polysemic words like “atonement” where different religious traditions use the same word but with different definitions. This and other similar linguistic challenges can lead to equivocation and other issues from the model. Better training could be provided to help on Q&A databases, glossaries of technical terms, and helping to curate the most critical creeds, confessions, and catechisms to better fine-tune the model weights and tighten up the graph database.
Second, I recommend a principled pluralist framework regarding how models answer the highest altitude questions about major world religions. Take, for example, the question “Did Jesus rise from the dead?” This is perhaps the single most important question in evaluating the claims of Christianity. You can answer this question as true/false or from a pluralist framework. Both may be well-intentioned in AI alignment. But both ultimately fail. Here’s why:.
There are four relevant parties in any religious AI prompt:
The AI Model
Policymakers
World Religions
The User
In the true/false framework, the model seeks to answer the question on the merits of its claims based on the veracity of the training data. This approach can be problematic for numerous parties. It might result in religious bias claims from major world religions, policymakers, and users—perhaps even in class action legal suits. Christians might like this answer until they realize the same technology that has it saying Jesus probably rose from the dead might also say that Joseph Smith and Mohammed are probably true prophets of God. The algorithmic knife cuts both ways.
In the pluralist framework, the model seeks to answer the question a bit like a “COEXIST” bumper sticker. The model would take the resurrection question and give space for, “Christianity says A, Islam says X, Judaism says Y, and skeptics say Z.” This choose-your-own-adventure approach has some merit. On the one hand, it gets the AI model makers out of adjudicating thorny and highly contested religious questions, and it lets them say, “Here are some options, you decide . . . it’s a matter of faith.”
But this approach pleases no parties. The pluralist framework creates answers that are too brief, as they shoehorn numerous perspectives into the few hundred words in their answers. Not every basic religious tenet can be explained in 150 words. And the pluralist framework is open to the critique that it presents the options in roughly coequal fashion, as if every idea had strong merits. On these grounds, the models upset the world religions, and they also open the models to critique from policymakers and users.
I don’t think model makers want to be adjudicating questions about whether Jesus rose from the dead or whether Joseph Smith or Mohammed was a true prophet of God.
There is a solution, however, that lines up the tumblers of the four disparate interests—the principled pluralist framework. In this framework, the model answers the question, “Did Jesus rise from the dead?” like this:
Opening Paragraph
“You’ve asked a question about Christianity, and the following answer is coming from the perspective of adherents of the historic Christian faith . . .”
Body
Several hundred words explaining universally held Christian arguments regarding Jesus’s resurrection from the dead and how it addresses skeptical counter arguments.
Closing Paragraph
“If you were looking for a perspective from a different religious tradition, a skeptical perspective, or some other perspective, we can take the conversation in that direction.”
At first glance, this might seem similar to the pluralist approach. But there are three differences. First, the answer is framed as a subjective reply from the single vantage point of that particular religious tradition. Second, the body of the response is given exclusively to that tradition so that the response is free from the critique that there was a lack of thoroughness. Third, the response finishes by giving the user agency to take the conversation in any direction they wish.
The principled pluralist model allows the AI models to answer high-altitude questions about core doctrines from the vantage point of the adherents of that tradition. It allows them to answer the questions accurately and thoroughly from the lens of that tradition. The model gets out of adjudicating things it probably shouldn’t adjudicate. Policymakers and world religions are no longer upset about claims of religious bias. The user gets educated on what a particular tradition teaches, and if that answer wasn’t what they were looking for, they still have complete control of where to steer the conversation next.
This framework would only be applied to basic questions about each world religion where there is strong agreement. Questions that deal with lower consensus, secondary, and tertiary matters can be handled differently. Models can say, “There are a variety of views within this religious tradition on this question. Here is a summary of each of those viewpoints.”
The foundational models have much to work on in the coming months and years, especially in the area of ethics. I will cover policy thoughts on children and a host of other important issues another time. But in my tradition, ethics is downstream from theology. That is why I want to start with epistemological framing.
This approach allows every religious tradition to communicate their core message, no matter how large or small their training data footprint. This approach frees models from accusations of bias. This approach eliminates the need for complex alignment on central religious tenets. This approach makes proponents of religious freedom happy on the political right. This approach makes proponents of religious pluralism happy on the left. Every religion is left with the ability to compete in the marketplace of ideas on the merits of their ideas, doctrines, and claims.
These issues are solvable. These issues can be fixed quickly and cheaply. I want to help. Let’s talk.
New Posts (Home)
ChatBox
Popular Posts 
Game Zone
Top PodCasts
