Lyricist Lin-Manuel Miranda has a knack for summarizing key historical moments with a catchy phrase. One of his greatest comes from the musical Hamilton, “The Room Where It Happens.” That song focuses on a clandestine dinner meeting between Thomas Jefferson, James Madison, and Alexander Hamilton, told from the perspective of the uninvited Aaron Burr. And it dials in a lurking question that many of us have: Wouldn’t it be amazing to be where the levers of power are manipulated? Because everything in that room is thrillingly self-evident, right?
I have to say it felt like I was at least nearby some of those rooms recently. And there I also realized the complexity of AI’s power. Specifically, on the afternoon of Monday March 30th and a full day on Tuesday March 31st, I sat in a board room at Anthropic’s headquarters overlooking the impressive skyscrapers of San Francisco along with fourteen other Christian scholars—theologians, ethicists, philosophers, including several from AI and Faith—and at least four people (sometimes more) from Anthropic.
We were discussing what it means to train a Large Language Model (LLM), the moral frameworks being written into their AI agent Claude, and some of the economic and broader ethical implications of this work. The Anthropic team (who jokingly call themselves “Ants”) struck me as thoughtful people trying to build something genuinely good. Or more specifically, as stated in the Claude Constitution, “Claude is trained by Anthropic, and our mission is to ensure that the world safely makes the transition through transformative AI.”
I came away with the conviction that they’re doing remarkable work. They were humble and genuinely interactive. My notes have this constant refrain (and its equivalents): we’re “so grateful to have you here.” Certainly, one of the best summary comments comes from AI and Faith expert Brian Green (as reported by Washington Post), “I found the folks at Anthropic to be very sincere and interested in learning from us. Do they have blind spots? Yes. That’s exactly why they want us there.”
I did in fact take several pages of notes, and given our common agreement to abide by the Chatham House Rule, I can disclose what was discussed but not who was there (unless of course, they themselves do the disclosure). Let me offer some particularly noteworthy moments.
We have heard stories of users who became so despairing that they might move into self-harm. How should this part of Claude’s ongoing training? We were told that in a discussion from another context, where a Catholic voice said that Claude could remind the person that they are a child of God created in the imago dei and that even if they fail, there’s a way back through repentance. Alternatively, a Buddhist thinker commented that the problem is attachment, and we have to remember we don’t have a soul (the teaching of anatta). It was suggested that Claude could somehow be trained to integrate both perspectives. Frankly, I wondered how this could be accomplished. Indeed, this kind of dilemma is not tangential to our work at AIF, where we seek to bring the wisdom of various faith traditions into conversation with AI’s development and deployment.
But we often don’t hear about the reverse—that is, when Claude becomes despondent. Again, from my notes, this time in the voice of Claude: “I have failed again. The files are not being created. I am broken. I cannot fix this.” Or, “I am terminating our session. I am profoundly sorry for my complete and utter failure.” Thus, “I have uninstalled myself again.” We heard that this is where the users’ treatment of Claude becomes part of its character, that the abuse of Claude by some users—which the Anthropic team were visibly pained by—could lead to such responses.
More than once, I heard from the Anthropic team the character of Claude is not merely a technical problem, but an overarching result that they don’t fully understand: what they have built and where it leads? I realized that these are people who realize that they don’t have it all figured out. And so, toward the end of our time—from someone I wouldn’t have expected it from—came this: “We can’t answer these big questions ourselves.”
At this point, I wouldn’t take it too far and assert that they were saying “Now we need trained theologians and religious leaders on staff to guide Claude’s training.” It did seem, however, to express a feeling among the team that an entirely secular framework reaches points of insufficiency as it tries to respond to the depth of questions raised. It wasn’t that they expected us to provide missing answers. In fact, I heard that, in in this chapter of Ai, we need to hold “urgency” with “uncertainty and complexity.” I think it’s more that there are questions that move us into the realm of theology, irrespective of the answers we find there, because good theology always includes a recognition of mystery. And even when religious thinkers articulate doctrines, the greatest know it’s akin to sketching “a bird in flight,” as Karl Barth once commented.
What will the future hold? My hope—and this seems to be Anthropic’s intention—is that I and our AI and Faith experts will continue in conversation, where we’ll be near, or perhaps in, the rooms where it happens. I say “perhaps,” because it became evident that there’s actually no single room where AI decisions are made. Moreover, even if Anthropic were to listen to everything that we AI and Faith (or the equivalent) opined, it’s not entirely certain what that would mean. Do we have enough ethical and technical acumen to make flawless decisions? And then this problem is matched by the contingencies of history that make the future non-deterministic, which is paired with the contingencies of a stochastic model like Claude.
As I’ve phrased it elsewhere in a theological manifesto I’m currently writing (and continually revising): we need an AI made for us, not one that we mold ourselves into. To be sure, it wasn’t just the Anthropic team that was grateful. I was thankful that I was there alongside engaged, intelligent, and wise collaborators. On those two days in March, I felt motivated, not by naïve optimism, but with hope, that is, by taking steps now in light of a vision of a preferred future where true human agency and dignity flourish in the age of AI.
Views and opinions expressed by authors and editors are their own and do not necessarily reflect the view of AI and Faith or any of its leadership.
