Sam Altman warns treating AI as religious authority is a real safety risk

Altman is drawing a bright line around one of the stranger failure modes of frontier AI adoption, which is not technical at all but psychological. If people start treating models as oracles with spiritual authority, you can get dangerous deference regardless of how careful the training was. His argument reframes model safety as partly a problem of human institutions and norms, not only alignment algorithms. That is a notable contrast with Anthropic’s willingness to engage religious scholars to think through ideas like model welfare.

This disagreement matters because it foreshadows different governance cultures as we inch toward more capable systems. One camp is experimenting with quasi moral frameworks for models; the other is warning that even talking that way could erode human responsibility. For regulators, ethicists and users, this is not an abstract dispute. It will shape whether future AI deployments come wrapped in language that invites deference, or in language that insists on human control and accountability.

From a competitive view, Altman’s stance also serves a strategic purpose. OpenAI is under scrutiny for its own safety culture, and pushing a message about preserving human judgment is a way of occupying the reasonable middle ground between AI cultism and blanket panic. As models get more persuasive and multimodal, the question of how society interprets their authority will sit alongside token pricing and context length as a core design choice.