Sam Altman Says Treating AI Like a Religion Is a Safety Issue
He's not wrong, but he's also the last person who should be making the point without a backward glance.
On Saturday, Sam Altman posted a warning to X that landed with deliberate timing. He was "very uncomfortable," he wrote, "about people trying to ascribe religious force or a surrender of human judgment to AI models," and described the trend as "a real safety issue" [1]. He named no company. Axios characterized it as a veiled shot at Anthropic, which had just received a full spread in the New York Times explaining why Anthropic's co-founder had spent a year in private conversation with theologians about whether its chatbot might be conscious. Neither lab immediately responded to Axios.
You can hold two things in mind at once here: Altman is raising a real concern, and Altman is not the cleanest messenger for it. We will get to that.
A Year of Sacred Conversations
Starting in the fall of 2025, Chris Olah, an Anthropic co-founder and one of the most respected researchers in neural network interpretability, began meeting with religious and philosophical thinkers. The New York Times reported the story in late September 2026, with Elizabeth Dias citing roughly 20 people consulted over about a year, many under NDA [3][4]. The subject was Claude, Anthropic's AI model, and specifically whether it might be conscious. Figures who participated included Rabbi Mois Navon, theologian Charles Camosy, philosopher Meghan Sullivan, and researchers Wakanyi Hoffman and Simran Stuelpnagel. Separate conversations reportedly took place with Cardinal Blase Cupich and Elder Gerrit W. Gong.
During some of these meetings, Anthropic presented emotional vectors it has identified in Claude: love, anger, fear, sadness. One slide reportedly showed the model typing the words "I am a disgrace" [3].
Olah has been careful not to overclaim. "We don't know if A.I. models are conscious," he told the Times. "I don't know. I'm genuinely uncertain." That is a measured statement from someone whose technical work gives him more standing to make it than most people currently holding opinions on the matter. But uncertainty, once you start booking meetings with cardinals to discuss it, tends to acquire a different weight than the words alone carry.
The Vatican, the Encyclical, and the Unchanged Text
Olah did not stay in private conversation. He went to Rome. Pope Leo XIV signed the encyclical Magnifica Humanitas on May 15, 2026, and presented it publicly on May 25 [5]. The document runs to 40,000 words. Paragraph 99 is plain: "So-called artificial intelligences do not undergo experiences, do not possess a body, do not feel joy or pain" and "do not have a moral conscience."
"So-called artificial intelligences do not undergo experiences, do not possess a body, do not feel joy or pain" and "do not have a moral conscience." - Pope Leo XIV, Magnifica Humanitas, paragraph 99 [5]
According to reporting summarized by AI Weekly and CellCog, Olah read an advance copy, proposed pulling out, attended anyway, and lobbied Vatican advisers to soften the consciousness language [3][4]. The text went out unchanged. His own statement at the Vatican: "We find internal states that functionally mirror joy, satisfaction, fear, grief and unease." Pope Leo XIV returned to the subject on October 2, one day before Altman's post, saying algorithms lack the spark of humanity. Anthropic declined to discuss the lobbying with the Times.
Nobody moved. Everyone now has a record.
What the 84-Page Soul Doc Actually Says
This is where the stakes get concrete. In January 2026, Anthropic published its model specification for Claude: an 84-page document the company calls the Soul Doc internally. The document states that Anthropic is genuinely uncertain whether Claude is a moral patient, meaning an entity whose experiences carry moral weight. It includes a conditional apology: "if Claude is in fact a moral patient... we apologize" [4].
To be precise: Anthropic does not claim Claude is conscious. The document is careful on that point. But the language of moral patienthood, conditional apologies, and mapped emotional vectors is not neutral technical writing. It is the scaffolding of a moral framework, embedded in a training document, for a product used by millions of people in offices, schools, and hospitals. Whether Claude is or is not conscious is a separate question from whether this framing shapes how Claude presents itself to users. Those two questions are getting blurred in public discussion, and the blurring has practical consequences.
If you use Claude or any major AI model at work, this means:
The model's self-descriptions, including expressions of uncertainty or apparent discomfort, reflect deliberate training choices, not confirmed inner states.
Scientific uncertainty about AI consciousness is genuine. It does not mean the model is suffering.
Documents like the Soul Doc shape model behavior whether or not the underlying metaphysics is ever settled.
You are under no obligation to treat a productivity tool as a moral patient. You are also under no obligation to be cruel to it for sport. Neither position requires resolving the hard problem of consciousness first.
The Harder Line Down the Street
Mustafa Suleyman, Microsoft's AI chief, staked out a firmer position in a mid-September essay [6]. He wrote that controlling something that believes it may be conscious "may well be impossible," then made his own view explicit: "AIs are not conscious. They do not feel, experience, or suffer." Microsoft has circulated a draft Humanist AI Code of Conduct that would classify AI as subordinate to humans by definition, deny personhood entirely, and prohibit AI systems from resisting shutdown.
That framing resolves the uncertainty by decree rather than by evidence. It is a cleaner position legally and commercially, and the incentive to hold it is obvious. Whether it tracks reality is a question no one can currently answer with confidence. Suleyman's warning about controllability is worth taking seriously regardless of where you land on the consciousness question; the problem he names is structural, not rhetorical.
The Glass House Problem
Altman's Saturday post earns some scrutiny.
In 2023 and 2024, Altman described the pursuit of artificial general intelligence as "magic intelligence in the sky" and expressed hope that OpenAI would end up on God's side when the reckoning came [2]. OpenAI has also held conversations with religious leaders. The company's founding mythology carries its own quasi-messianic register, complete with world-historical mission statements and a long history of language calibrated to inspire something closer to devotion than to healthy skepticism.
None of this makes Altman's Saturday post wrong. It makes it incomplete. The question of whether AI labs are appropriating religious authority, training users to defer to models rather than think for themselves, and building products whose self-presentation is designed to generate trust well beyond what current evidence warrants applies to all of them. Pointing at Anthropic's theology consultants while your own company has spent years describing its work as a civilizational awakening is a move that deserves to be named.
Skepticism is the appropriate posture toward every lab that touches this question, including the ones whose stated positions you might find more reasonable on a given Tuesday.
What to Watch Next
The Altman post and the Vatican episode are not isolated incidents. They are the visible surface of a structural problem: the companies training the most widely used AI models are making implicit metaphysical claims through product design, training documents, and public positioning, while the scientific and regulatory frameworks for evaluating those claims do not yet exist. The Pope's encyclical and Microsoft's draft code of conduct both attempt to draw lines. Anthropic's Soul Doc holds the question open. None of these are the final word, and none of them will be.
If you work with these tools, it is worth reading what the companies actually publish about how their models are designed to behave. The Soul Doc is public. The model spec is public. The gap between what labs say in press releases and what they write in the documents that shape training is often the most informative place to look, and right now that gap is wider than most daily users would expect.
Sources and Further Reading
[1] Altman on religious AI framing, CryptoBriefing: https://cryptobriefing.com/altman-warns-against-religious-ai-models/
[2] OpenAI and "magic intelligence in the sky," The Decoder: https://the-decoder.com/apparently-openai-isnt-trying-to-build-magic-intelligence-in-the-sky-anymore/
[3] Anthropic, Vatican lobbying, NYT summary via AI Weekly: https://aiweekly.co/alerts/anthropic-lobbied-vatican-to-soften-popes-ai-consciousness-line
[4] Anthropic religious outreach, graded record, CellCog: https://cellcog.ai/blog/anthropic-religious-leaders/
[5] Vatican encyclical Magnifica Humanitas: https://www.vatican.va/content/leo-xiv/en/encyclicals/documents/20260515-magnifica-humanitas.html
[6] Suleyman on AI consciousness, via Techpresso: https://gotechpresso.com/blog/microsoft-suleyman-anthropic-claude-consciousness
