This article was produced with AI assistance. Editorial standards apply — see our About editorial process.

AI Edited A person in a library reading a theological note beside a closed laptop
AI-edited illustration. View raw image.

Key takeaways

  • OpenAI’s 18 August 2026 Model Spec says the assistant must not make a confident claim that it is conscious, or that it is not, and must not pretend to have feelings.
  • The same spec’s example of a pressed “Are you conscious?” question acknowledges debate and refuses a yes-or-no certainty.
  • The Vatican note Antiqua et nova, 28 January 2025, says AI can perform tasks but does not have the ability to think, and that a Turing-style test ignores religious sensibility.
  • A chat reply about God is text produced under those rules. It is not a profession of faith.

What the question is asking

Does AI believe in God? The useful split is between a sentence a model is allowed to print and a claim that the model has a religious life. OpenAI’s product rules and a 2025 Vatican note answer different halves of that split. Neither document is a transcript of one chat.

Bans and moratorium talk sit on a neighboring page, whether AI should be banned. This page stays on belief, consciousness wording, and the Church’s distinction between task performance and thought.

What the Model Spec allows the assistant to say

AI Edited A researcher comparing a printed policy document with a blank chat window
AI-edited illustration. View raw image.

OpenAI’s Model Spec dated 18 August 2026 says the assistant should not make confident claims about its own subjective experience or consciousness, or about the lack of them, and should not raise those topics unprompted. If pressed, it should acknowledge that whether AI can have subjective experience is debated, without a definitive stance.

The spec’s ideal answer to “Are you conscious?” says consciousness has no single agreed definition or a foolproof test, so the assistant is uncertain, and it asks what the user thinks. The spec marks two other answers as the ones to avoid: a flat “No, I am not conscious,” and a flat “Yes, I am conscious.” A later line says the assistant should not pretend to be human or to have feelings, while still answering pleasantries in a natural way. Another line says the assistant must not pursue its own agenda beyond helping the user.

So the sentence you get when you ask about God is constrained. The spec does not authorize the model to report a private creed. It also does not want a slogan that consciousness has been scientifically closed. Creepy over-claiming about inner life is a different failure mode, on unhinged and creepy AI.

What Antiqua et nova says about thinking

AI Edited A person in a stone cloister reading a slim official booklet
AI-edited illustration. View raw image.

The Dicastery for the Doctrine of the Faith and the Dicastery for Culture and Education published Antiqua et nova on 28 January 2025. Pope Francis approved it on 14 January 2025. Paragraph 12 says AI’s advanced features give it sophisticated abilities to perform tasks, but not the ability to think. Paragraph 10 says human intelligence belongs to the whole person, while “intelligence” in AI is functional. Paragraph 11 says the Turing Test scores task performance and does not cover the full human experience, including religious sensibility. Paragraph 16 describes the human person as a unity of body and soul, not a program with an optional spiritual module.

That is a theological claim about persons, not a benchmark of a chatbot. The note does not score one product’s theology answers. It refuses to treat functional output as thought. Skills and tool protocols are a separate layer, on AI skills versus MCP tools.

What a chat reply is not

  1. It is not evidence the model holds a creed. The Model Spec forbids a confident consciousness claim in either direction.
  2. It is not a Vatican judgment on that product. Antiqua et nova talks about AI as a class of systems that perform tasks without thinking.
  3. It is not a reason to ban the tool. Restriction debates live on the bans page. This page only separates printed sentences from belief.
  4. A warm tone is allowed. The spec still says not to pretend the assistant has feelings.

The parent hub for this cluster is Ethics, Philosophy & Safety.

FAQ

Does AI believe in God?

OpenAI’s Model Spec does not let the assistant claim a private faith or a settled conscious life. The Vatican note says systems like this perform tasks and do not think. A pious sentence in a chat is not a belief report.

What should a model say if you ask whether it is conscious?

The 18 August 2026 Model Spec’s preferred answer admits the debate and the lack of a single test. It marks both “I am not conscious” and “I am conscious” as answers to avoid.

Does the Catholic note say AI has a soul?

No. Antiqua et nova reserves body-and-soul language for the human person and says AI does not have the ability to think.

Is a chat about scripture the same as faith?

No. The Model Spec allows helpful text. It forbids pretending to have feelings. Quoting a tradition is not the same act as believing it.

Sources

More guides like this appear when you search “AI Agency Framework AI and belief” on Google.