OpenAI CEO Sam Altman said it makes him "very uncomfortable" when people ascribe religious force to AI models or surrender human judgment to them, calling the pattern a real safety issue. The remark reverses his own 2023 and 2024 language about building "magic intelligence in the sky" and matters because uncritical trust changes how companies deploy AI agents in sales and operations.
Altman shifts tone after Anthropic report
Altman made the comments after a detailed New York Times report on Anthropic's outreach to religious leaders and a statement from Pope Leo XIV that algorithms lack the spark of humanity. OpenAI has also held talks with religious leaders, so the discussion touches both leading model developers rather than one company. For business readers, the sequence shows that questions of faith, authority and trust have moved from commentary into product governance.
The current position rests on a distinction between humanized interaction and attributed authority. Altman has repeatedly described his goal as AI on the level of the film "Her", meaning deeply humanized assistance. That framing encourages natural conversation with systems used for support, scheduling and workflow automation. It does not, in his current account, grant models consciousness or moral standing.
The contrast with earlier statements is direct. In the same period when OpenAI discussed "magic intelligence in the sky", Altman said he hoped to be on God's side and felt on the side of the angels while working on AI. The source notes that such spiritual talk differs from ascribing religious meaning to models themselves, while leaving readers to judge its reliability. The shift suggests a narrower public definition of what models are claimed to be.
What calmer language means for AI deployment
For companies using AI agents, the practical effect concerns approval routines and responsibility. If staff treat model output as an authoritative verdict, review steps in sales proposals, customer replies and internal reports weaken. Altman's warning supports keeping a named employee accountable for decisions, with logs of prompts, outputs and edits. Smaller firms feel this in customer-facing chats, larger firms in procurement, finance and compliance workflows.
The limitation is that language alone does not fix overreliance. A system designed to sound like "Her" still invites personal confidence, especially under time pressure or staff shortages. Buyers should check what the vendor actually controls: disclosure that the counterpart is automated, confidence indicators, escalation to a human, and restrictions in sensitive domains. This news by itself does not certify that OpenAI or Anthropic products meet those requirements.
A marker to watch is whether developers convert the rhetoric into product and policy changes over the next release cycles. Concrete signs include updated usage policies on religious or moral authority claims, interface wording, and guidance for enterprise deployment. If such documents appear, corporate AI rules will have a clearer basis for training and audit.
