Anthropic's theologians prompt Altman's warning against deifying AI
A quiet experiment inside Anthropic — bringing religious scholars into the work of building Claude's morals — has become the industry's newest fault line, with Sam Altman weighing in from the sidelines.
Anthropic has spent nearly a year privately convening theologians to help shape how morality gets built into Claude, and Sam Altman says treating AI as a spiritual object is itself a safety problem. The talks, begun in fall 2025 and led by Anthropic co-founder Christopher Olah, gathered scholars from Catholic, evangelical, Jewish, Sikh, Hindu and Vedanta traditions to work through questions like how a model should reason about right and wrong — and whether models could ever be conscious or hold moral status. The sessions fed an 84-page internal draft constitution for Claude, nicknamed the "Soul Doc," according to the New York Times, which first reported the consultations. One Hindu monk, Swami Sarvapriyananda, flew to San Francisco under non-disclosure agreements for what Anthropic called "wisdom tradition circles" and says he spent "hours and hours" with a founder; Olah also reportedly lobbied the pope's advisers before a May Vatican event to take conscious models seriously.
Altman broke cover on Saturday with a post on X that stopped short of naming Anthropic but left little doubt about the target. "I am very uncomfortable about people trying to ascribe religious force or a surrender of human judgment to AI models, and think it is a real safety issue," he wrote. Google DeepMind's Jon Barron drew the same line more bluntly, saying he is "disturbed by how some of my fellow AI researchers keep trying to elevate the moral standing of checkpoints and harnesses."
The split matters because it is no longer philosophical navel-gazing — it is a fight over what counts as safety work. Anthropic's position treats a model's possible moral status as a design input: if Claude might one day matter morally, you want its values drafted carefully, by people who think about values for a living. The OpenAI and DeepMind position is that this kind of reasoning encourages users (and maybe the labs themselves) to over-attribute judgment to software, which is how a safety culture turns into a theology. Both camps claim the safety high ground, and the answer will shape how much of Claude's constitution stays behind NDA versus out in the open. The title "Soul Doc" being public at all is a sign that Anthropic's experiments are leaking faster than its explanations.
Should labs be required to publish the constitutions their models are trained on — soul documents included? Tell us in the comments.
