OpenAI CEO Sam Altman has cautioned against treating AI models as possessing religious or magical properties, describing such attribution as a "real safety issue." His warning comes amid growing engagement between AI companies and religious institutions, including reported meetings between Anthropic and religious thinkers, as well as recent papal commentary on artificial intelligence.
Altman's current stance represents a notable shift from his own earlier rhetoric. As recently as 2024, he publicly spoke of building "magic intelligence in the sky" and characterized his work on AI as feeling "on the side of the angels." Those statements reflected language that positioned AI development in quasi-spiritual terms.
The CEO's pivot signals heightened concern within OpenAI about how the public and religious communities frame AI capabilities and purpose. Attributing supernatural or divine characteristics to AI systems carries concrete risks beyond philosophical debate. If users or institutions treat advanced models as infallible or possessing inherent wisdom tied to moral authority, it could drive misplaced trust and poor decision-making in critical domains.
Altman's framing of religious attribution as a safety issue aligns with broader AI safety discussions around alignment and responsible deployment. Systems trained on human data reflect human biases and limitations. Enshrining them with religious significance could amplify those flaws while obscuring their actual constraints and failure modes.
The timing of Altman's comments reflects industry attention to how mainstream institutions interpret AI. Anthropic's engagement with religious thinkers suggests deliberate outreach to faith communities, potentially to shape how religious groups understand AI development and use. Such conversations carry weight given religious institutions' influence over large populations and their role in ethical frameworks.
Altman's current messaging appears designed to establish clearer boundaries between technical capability and transcendent claims. AI models excel at pattern recognition and language generation within specific domains. They do not possess consciousness, moral agency, or access to truth beyond their training data