Why Sam Altman Says Giving AI Religious Authority Is a Safety Risk

Sam Altman said he was “very uncomfortable” with people ascribing “religious force” to AI models or surrendering human judgment to them, calling it “a real safety issue.” He did not name Anthropic in the post, but its timing—days after reports about the company’s meetings with religious scholars—led coverage to read it as a veiled criticism. 71

The distinction at the heart of the dispute matters: asking how an AI should behave, or whether it could be conscious, is not the same as telling people to treat it as a spiritual authority. The available reporting describes Anthropic exploring those questions, while its co-founder Christopher Olah said he was genuinely uncertain about whether AI models are conscious. 1

What Altman warned against

Altman’s post joined two concerns: treating AI as if it held religious power and handing over human judgment to a model. He called that combination a safety issue. Because the post did not explicitly identify a company or a project, the connection to Anthropic is an interpretation based on the surrounding news, not a direct accusation in the post. 7

That caution is broader than a debate about religion. When people treat a model’s confident, human-like answers as a source of authority, they may give it more influence over decisions than its demonstrated understanding warrants. The cited coverage documents Altman’s warning; it does not establish how widely people are doing this or measure its effects. 67

Why Anthropic’s meetings drew attention

Anthropic consulted religious scholars about how to shape Claude’s moral behavior and whether AI models might be conscious. Those topics overlapped with Altman’s warning, which is why his remarks were read as a response to the company’s work. 17

But the distinction is important. Olah said, “we don’t know if A.I. models are conscious,” adding that he was genuinely uncertain. Discussing possible consciousness is not proof that a model has subjective experience, and consulting religious thinkers about moral behavior does not by itself give a model religious status. 1

Consciousness and moral behavior are separate questions

A model can be designed to follow guidance about how it should respond without that guidance establishing that the model has an inner life. The reported Anthropic discussions touched both Claude’s behavior and possible consciousness; those questions should not be collapsed into one claim. 1

The reporting also records objections to treating models as conscious in training: a Microsoft AI executive warned that doing so could be dangerous. That is a reported concern, not a resolution of the wider question of machine consciousness. Olah’s own stated position, as reported, was uncertainty. 91

The human-centered stakes

Pope Leo XIV’s AI encyclical put human dignity at the center of the discussion. Reporting on the document says it called for stronger oversight of AI companies, protections for workers facing disruption, and measures against fake AI-generated information. These concerns reinforce the importance of keeping human interests and accountability central as AI systems are developed and used. 152

OpenAI faces its own scrutiny

The debate unfolded amid separate reports about OpenAI. Business Insider reported that safety leader David Robinson left the company and criticized its safety record; the BBC reported that OpenAI fired three researchers over alleged mishandling of information. These reports do not answer whether AI models can be conscious, but they are relevant context when comparing the companies’ claims and governance. 1226

The clearest takeaway is narrower than a verdict on Anthropic or OpenAI: Altman warned against treating models as religious authorities or yielding human judgment to them. Anthropic’s reported consultations raised related questions about Claude’s moral behavior and possible consciousness, but the available reporting does not establish that those discussions amounted to giving Claude religious authority—or that AI consciousness has been settled. 71