Weekend Round-Up

More on Magisterium AI, Martyrdom and the Unborn, and Archbishop Fernández on Theology

In lieu of a full article, here is a weekend round-up of reflections on Catholic news.

(Photo: Sincerely Media, Unsplash)

---

Further Thoughts on Magisterium AI

Earlier this week I wrote about my experience using Magisterium AI, the large language model (LLM) designed to answer questions about Catholic doctrine and theology. Although the program is good at answering questions about doctrine, I found that it struggles with a historical perspective, often failing to correctly describe the views of historical figures like St. Augustine or St. Thomas Aquinas, and not being able to describe change in Catholic teaching over time.

I also found that Magisterium AI could be prompted to do other things besides answer doctrinal questions. Although not very good at writing homilies, it could, with some trial and error, offer interesting insights into biblical passages. I also managed to “jailbreak” the AI, that is, convince it to do something it was designed not to do. For example, if an LLM was programmed not to provide instructions on how to build a bomb, a creative hacker could potentially get around this by asking the LLM to write a story in which one character tells another how to build a bomb. My jailbreak was, by comparison, pretty benign. I asked the LLM to give me some advice on how to deal with the difficulties of going through a career change, to which it replied that such questions are outside of its purview. By asking the LLM what advice a priest could give to someone in that situation, however, I was able to get an answer. I confirmed this loophole with three other examples.

After I published that post, I continued experimenting and found a simpler solution. If, instead of asking for “advice,” I asked Magisterium AI for “spiritual advice,” for many prompts it was willing to give me an answer. I found this pretty amusing. Imagine a parishioner coming to visit the priest at the parish offices and asking for advice, only to be turned away since it is outside the scope of the pastor’s duties; a couple of seconds later, the parishioner turns around, goes back into the pastor’s office and asks: “Could you give me some spiritual advice?” “Well, in that case…”

The reason this is funny is because we recognize that it is what the person wants advice about that really determines the content of the question, not whether they call it “advice,” “spiritual advice,” or something else. Magisterium AI, on the other hand, needed a verbal marker to recognize that the question is something that it can handle. But this difference is also theologically interesting. As human beings, we instinctively intuit that a question about struggling with career change, the loss of a loved one, or something similar, has a spiritual dimension. Magisterium AI, on the other hand, needs to be told to think about the question in spiritual terms. It operates out of a kind of theological extrinsicism. One exception is questions concerning morality, particularly issues that are discussed in the Catechism or other recent Church documents, which it seems to recognize as within its ambit.

Oddly, asking Magisterium AI for “spiritual advice” did not work as often as asking what advice a priest should give on my topic of concern. Also, it did not get me the same results. Asking for “spiritual advice” usually resulted in a lengthier, more didactic treatment of the topic I asked about, while asking what advice a priest should give gave me more direct, and more helpful, answers. I don’t know what to make of that.