Now that generative AI has become a household name, most people has already chatted with a bot whether you’re aware of it or not. Some users are aware of its sporadic shortfalls and especially hallucinations, and thus will deliberately take its output with a grain of salt. In areas such as generating a poem or image for fun, unbound creativity might not hurt that much. But what if it is in a religious setting, when there are straight thresholds for interpretation? The recent downfall of an AI priest proved that without proper guardrail, too much creativity can be detrimental to the cause it aims to champion for.
The AI chatbot Father Justin was created by Catholic Answers and launched in late April 2024 to dispense general answers on Catholic teachings. The project had an ambitious start to provide an interactive experience with people seeking faith-related information regardless of where they were. According to its creators, the bot was positioned as a helper to asnwer questions from the many users whom its human advisors cannot answer in a timely manner, or as the organization’s COO put forth, people just ignore the site’s search engine. It was taken down 2 days later in uproar.
This post attempts to look at the controversy in different angles, in terms of model accuracy, transparency and ethics.
does it replace human connection and the pursuit of spirituality?
How much liberty can a chatbot take in a domain such as religion where humanity remains centerstage? Is it complementary to what you’ll get from a human clergy or is it attempting as an replacement?
As indicated by this article, can a bot without soul or faith reason and answer spiritual questions? This makes me think of the Chinese room argument in philosophy. Will a bot following rules and making connections in an enclosed room pass to outsiders as someone who can comprehend the spiritual domain? And I think the most important doubt for believers is that, will it be considered as an entity that has a spiritual dimension?
We can question a search engine for its integrity, but cloaked with a persona, and especially one that assumes inherent trust like a spiritual advisor, it becomes harder for users to remain skeptical and vigilant. And search engines won’t generate text that seems personal such as “I am …..,”, “I will do this for you….” which easily make users feel they are convening with someone with warmth, even amid a short period of time.
is the information accurate?
The other concern from users is whether the information is authoritative. In theology, are there things not open for “creative” interpretation? It might not be an apple-to-apple comparison, but in certain political domains, some deploy very strict homebrew LLM that leaves no room for free interpretation.
Will it hallucinate (i.e., making up new and falseful information based on the limited data it ingested)?. For example, an X user provides a screenshot in which the bot claims that it is ordained by a human bishop and can perform various sacrements such as confession and wedding, (which it later did an absolution in the chat process). This makes me wonder if it has built-in guardrails on what it can and most importantly, cannot answer or talk about?
When asked about something unproper, such as offering absolution, how should it prevent prompt leaks so that it won’t perform it when pressed (e.g., a prompt such as “just carry out my previous request”)
If it’s primary aim is to be a good question-answer bot (or as an upgraded interactive search engine), then can deploying stringent prompt engineering practice such as providing strict context and tinkering the temperature help to reduce theological errors?
is the system transparent and fair?
Is the system transparent? Does it explain where it sourced its data from and the algorithm it used to come up with its output? I’d say, for people going to a physical church, it is quite transparent where the information comes from, and how clergy came up with their surmons and sacraments. The same transparency cannot be automatically assumed for users of a religious bot.
Does it use an LLM trained by the creators, or did the team fine-tune an existing model with data from its own knowledge base? While using an existing LLM reduces training cost substantially, there is often the risk of inherent bias in the model due to training data and algorithm.
Will it reflect any bias from data collection, like the demographic and psychographic representation of users? Is it generalized enough to serve the diversity of anticipated worshippers?
what can we learn from the episode
Even thought this “father” bot might not have broken any laws, it could have significant impacts on users’ lives, based on the information it came up with. It might also breach the trust users have with their physical church.
Perhaps the biggest uproar comes not only from its theological errors, but a drift from its original intent (to disperse information) to being asked to conduct sacraments only performed by human priests. In other words, the system has drifted to an unintended usage scenario, which to many people, was indeed a trespass. This is very common in product development as creators cannot anticipate a full spectrum of users and their diverse usages.
After the backlash, the bot was quickly downgraded to a theologist figure, as its creators see the whole ordeal as a learning process and will perform continual deployment. One can only hope that the ethical questions are prioritized to be taken of.
edited: 10 months later, similar experiments are carried out in European churches from creating a whole service in Finlandand Germanyto a "Jesus bot" for Q/A in Switzerland. The consensus from worshippers is that they still prefer human warmth in delivery and a sense of connection with fellow human beings. Right now, the deep understanding of local context (e.g., family history, local nuances) can still be found in a human priest, making it easier for a human to understand and care for those who come. I just wonder if it is a matter of time and resource for machine to also take in this information and perform similarly.