Sam Altman’s post on X on Saturday did not announce a product, a funding round, or a shipping date. It raised a question the artificial-intelligence industry rarely handles in public: whether people are beginning to treat AI models the way they treat religion. Treating a model as though it carries “religious power,” Altman wrote, is a real safety problem.
The post named no company. Axios read it as a comment aimed at Anthropic, the rival lab that has spent more than a year meeting with religious figures to discuss where its Claude model belongs, ethically speaking. OpenAI, which Altman runs, has taken a different line, and the two companies now sit on opposite sides of a debate that is less about theology than about who gets to define what a frontier model is.
The argument is not about whether anyone literally worships a chatbot. It is about authority. Altman’s concern, as researchers following the exchange described it, is that deference is the danger. A model trusted the way people trust a religious authority stops being checked the way a machine is supposed to be checked. When an output is received as doctrine, a wrong answer does not get corrected; it gets believed.
Safety researchers have long listed over-trust as one of the quieter ways an AI system fails. A model that is wrong but questioned causes a limited amount of damage. A model that is wrong and treated as authoritative causes more, because its errors travel further before anyone stops to look. That is the distinction Altman was drawing, and it is why a word like “religious” lands as a safety claim rather than a culture-war aside.
Anthropic has spent the past year working the other side of the same question. The company has held discussions with religious leaders and ethicists about Claude’s ethical positioning, part of a long effort to build its brand around safety as much as capability. The conversations have covered what a model is allowed to say, what it should refuse, and what role, if any, a machine can play in questions of meaning. Where Altman sees a risk in religious framing, Anthropic has treated it as a subject worth engaging directly rather than avoiding.
That difference reflects two different companies. Anthropic has built its reputation on the safety of its models as much as on their performance, and its outreach to religious figures is consistent with that identity. OpenAI has pushed harder on scale and speed, and its chief executive has spent years warning that advanced AI could be dangerous. The religious question exposes a fault line that has been there for a while, buried under a shared enthusiasm for building the models in the first place.
The question reached Rome this week. In his first encyclical, Pope Leo XIV declined to give AI any kind of personhood. Machines, the document argued, cannot replace human judgment, and no matter how capable a system becomes, it does not carry the kind of standing that belongs to a person. The encyclical was read in the industry as a firm answer: a model, however powerful, cannot hold anything like religious authority.
The leaders of Google DeepMind and Microsoft’s AI efforts have also joined the discussion, according to people familiar with the matter. Their entry is a sign that the question has moved beyond the two companies that started it. What began as a barb between rivals has become a debate the whole field now has to answer, because the language used to describe these systems shapes how they are regulated, priced, and trusted.
The timing matters. Washington has been replacing “artificial intelligence” with “superintelligence” in official documents, and companies have followed. Names carry expectations. A tool can be questioned; an oracle cannot. If the industry settles on language that elevates models above ordinary software, it inherits the burden of explaining why a system granted that standing should ever be doubted.
There is a business question buried in the theology. Both OpenAI and Anthropic are trying to sell models to enterprises that need to be able to explain every answer a system gives. A model positioned as an authority undercuts that promise; a model positioned as a tool supports it. The companies’ different approaches to the religious question are, in part, different answers to the same commercial problem: how to be trusted by customers who will fire them the moment an output cannot be defended.
For Anthropic, the episode cuts closer to home. The company has spent years positioning Claude as the model that thinks about consequences, and its religious outreach is part of that. Being singled out by the head of a rival, however obliquely, forces a question the company has mostly answered in private: whether engaging religious institutions strengthens its safety brand or invites the charge that it is dressing a product up as something sacred.
For now, the debate is a war of framing. Altman posted; Anthropic keeps talking; the Vatican has weighed in. None of it settles the underlying question, which is whether a machine that can out-reason most people should be treated as something more than a machine. The answer will not come from a single post or a single encyclical. It will come from what people actually do the next time a model tells them something they do not want to hear.
