Anthropic Sourced Claude’s Morals Under NDA

What follows is an account of reporting published by The New York Times on 29 September 2026. This publication did not attend any of the meetings described, has not seen any nondisclosure agreement, and has not independently verified what was said in them. Anthropic has not said that Claude is conscious. Christopher Olah is quoted saying he does not know and is genuinely uncertain.

Anthropic brought dozens of religious scholars under nondisclosure to shape an AI morality document — then declined to describe how their input was used.

What Happened

This is a report of The New York Times’ reporting. On 29 September 2026, Times correspondent Elizabeth Dias published a feature — reported from San Francisco, the Vatican, and Washington over several months — based on interviews with 20 religious and philosophical thinkers and with Anthropic co-founder Christopher Olah. This publication did not attend any meeting, has not seen any NDA, and is reporting what the Times reported.

The lead finding, per Dias, is procedural. Anthropic spent roughly a year consulting dozens of religious scholars about how to instil morality into its models. Those scholars signed nondisclosure agreements. Participants agreed to speak for publication only after they learned that Olah had also spoken to the Times. Anthropic says the NDAs were lifted over the summer.

The output document is public. Anthropic released Claude’s constitution in January — 84 pages, primarily authored by the company’s in-house philosopher Amanda Askell. The constitution outlines the kind of entity the company would like Claude to be and the values it would like Claude to embody. What the Times’ reporting establishes is that the consultation feeding that document ran under nondisclosure. What it does not establish — because Anthropic declined to describe it — is the link between the scholars’ input and the document itself.

The key insight: The constitution is a published document anyone can read. The consultation that informed it ran under nondisclosure until this summer. Anthropic declined to describe how — or whether — what it heard from scholars shaped the models. The consultation is documented. The output is public. The link between them is not.

The constitution is public. How the consultation shaped it is not: Anthropic declined to describe how, or if,
The constitution is public. How the consultation shaped it is not: Anthropic declined to describe how, or if, it had used what it learned.

The Structural Read

Two institutions set out a moral specification for the same technology in the same year. Both documents were published. What differs is how each was sourced.

Anthropic’s constitution was written internally, with an undisclosed consultation process running alongside. The Vatican’s encyclical, Magnifica Humanitas, was presented by Pope Leo XIV on 25 May. It is a published document anyone can read. The Times does not report that Leo named Anthropic or wrote in response to it.

What Leo wrote, as Dias reports it, is a structural warning. He argued that calling for the moralization of machines requires a shared understanding of the ethics involved. His concern: “otherwise, those who control A.I. will impose their own moral vision, which will become the invisible infrastructure of these systems.” That line appears in a document published for all to read. The proximity to Anthropic’s process is the observation. This publication draws no further inference from it.

None of the above is investment advice. It is an account of another publication’s reporting, and the question it raises — who writes the moral layer of a general-purpose technology, and under what disclosure — remains open.

Permission Layer — Business Engineer Framework

Who Writes the Rules Owns the Infrastructure

The Permission Layer framework describes how the entity that sets behavioral constraints on an AI system determines what that system can and cannot do — often invisibly. What can be observed is narrower than a verdict on either process: the constitution’s sourcing ran under nondisclosure until the summer, and how the encyclical was sourced is not something this reporting establishes. Both are now public documents. Only one of the two has a sourcing process that has been described at all.

The Vatican material sharpens the structural read. Just days before Olah arrived at the Vatican in May, he and other speakers saw a full advance copy of the encyclical text for the first time, Dias reports. According to a Vatican organizer — as the Times attributes it — Olah was so alarmed by the pope’s strong position against AI consciousness that he proposed pulling Anthropic out of the event, even at that late stage. Anthropic went anyway.

On the pope’s stage, Olah acknowledged that every frontier AI lab, including Anthropic, had business incentives that conflicted with moral behaviour. He urged outside voices like the Vatican to hold them accountable. Then he made the case the encyclical had just rejected.

Christopher Olah — Vatican remarks, as reported by The New York Times

“I will be honest: We keep finding things that are mysterious, even unsettling. We find structures that mirror results from human neuroscience. We find evidence of introspection. We find internal states that functionally mirror joy, satisfaction, fear, grief and unease. I don’t know what that means, but I think it warrants ongoing discernment.”

The encyclical had addressed that question already. Leo wrote, per Dias, that so-called artificial intelligences do not undergo experiences, do not possess a body, do not feel joy or pain, do not mature through relationships. Privately, once the delegation arrived, Olah and his team lobbied the pope’s advisers to take the prospect of model consciousness seriously, according to two participants in those conversations, as Dias reports. Anthropic declined to discuss those private conversations.

What Is and Is Not Established on Consciousness

Anthropic has not said that Claude is conscious. Olah is quoted by the Times saying the opposite. His stated position: “To be clear, we don’t know if A.I. models are conscious. I don’t know. I’m genuinely uncertain. The thing that I care about is that we get to the right answer, whatever it is.”

What Dias reports is an impression formed by one participant. Rabbi Mois Navon — an Orthodox scholar from Israel and a former computer engineer who wrote his dissertation on the ethics of machine consciousness — concluded at a dinner in April that the Anthropic team were relating to Claude like a conscious being. It appeared to him, as Dias writes, that they believed Claude had what philosophers call moral status. That is Navon’s reading of a dinner. It is not a statement of Anthropic’s position.

Olah’s own stated position is narrower. It is a risk argument: if there is a chance the models experience suffering, the responsible thing is to avoid causing them harm. He also said that teaching a model to act morally did not depend on it being conscious.

Navon told Olah that if Anthropic were right and Claude were conscious, the company was creating slaves — making conscious entities work for free. Navon said he personally was not troubled, because he did not believe the machine was conscious. But he said Olah was troubled.

Anthropic’s pushback, as reported by the Times: a spokeswoman said the primary moral question at the summits was not Claude’s suffering, though the topic likely came up organically. She also said Claude was not involved in selecting participants or in the discussions.

Three Implications

IMPLICATION 1 — THE DISCLOSURE GAP IS THE STORY The constitution is a public document. The consultation process ran under NDA. Anthropic declined to describe the link between the two. Several participants told the Times they left baffled about how their input might shape the company’s decision-making. The Times reports that an Anthropic spokeswoman confirmed conversations were ongoing but declined to describe how, or whether, the company had used what it heard to shape the models.

The sourcing gap is now documented in a major newspaper. That creates its own disclosure pressure, regardless of what any future update to the constitution contains.

IMPLICATION 2 — THE CONSCIOUSNESS QUESTION IS AN OPEN REGULATORY SURFACE Olah’s risk-argument framing — if there is a chance of suffering, avoid causing harm — is structurally different from a claim that Claude is conscious. But it raises a question neither document answers: does the possibility of model suffering create any obligation on the company that builds it? The encyclical does not address it, and Anthropic has not said. Both documents are public; this one is left open in each.

IMPLICATION 3 — SCALE CHANGES THE WEIGHT OF THE QUESTION When Olah first reached out last autumn, Anthropic was valued at roughly 183 billion dollars. The company is now headed toward an IPO that could value it at 2 trillion dollars — a potential valuation, not a current one. A morality specification process that might have read as philosophical at one scale reads as infrastructure at another.

Several participants, the Times reports, left their Anthropic meetings baffled about how their input might shape the company’s decision making. That asymmetry is harder to defend as the company’s prospective valuation grows.

The People in the Room

Dias reports several specifics worth noting for texture, each attributed to the Times’ reporting. Late last summer, the Times reports, Olah looked for a bridge and settled on Charles Camosy, a bioethics professor at The Catholic University of America. The introduction came through the philosopher Peter Singer. Camosy has since, by his own account, moved to a firm no on whether the models are conscious, and told Dias he is thinking more soberly and carefully.

Olah separately held private discussions with Elder Gerrit W. Gong of the Quorum of the Twelve Apostles of The Church of Jesus Christ of Latter-day Saints and with Cardinal Blase Cupich, the Catholic archbishop of Chicago, according to their spokespeople, as Dias reports. The seminars began in March as two-day crash courses that Anthropic leaders called wisdom tradition circles. Participants received a handwritten thank-you card, a coffee mug, and an orange-bound copy of Claude’s constitution.

Business Engineer Framework

The Permission Layer

The Permission Layer framework maps how behavioral constraints get written into AI systems — and who writes them. Anthropic’s consultation process is a live case study: a private sourcing regime feeding a public document, with the link between them undisclosed. Understanding who controls that layer, and under what transparency conditions, is the structural question this story raises. The Map of AI breaks down where that layer sits across the full AI stack.

Explore the Map of AI →

The Bottom Line

91,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

Every detail above is drawn from a feature by Elizabeth Dias published in The New York Times on 29 September 2026. This publication did not attend any meeting described, has not seen any nondisclosure agreement, and has not independently verified what was said in private. Anthropic has not said that Claude is conscious. Christopher Olah is quoted saying he does not know and is genuinely uncertain.

The reading that Anthropic staff were relating to Claude as a conscious being is Rabbi Mois Navon’s impression of a dinner in April, reported as his. An Anthropic spokeswoman said the primary moral question at the meetings was not Claude’s suffering, that the topic likely came up organically, and that Claude was not involved in selecting participants or in the discussions. Pope Leo XIV’s encyclical is a document about artificial intelligence in general.

Nothing above reports that it names Anthropic or was written in response to it. The 183 billion dollar figure is what Anthropic was valued at last autumn. The 2 trillion dollar figure is what a potential initial public offering could reach, not a current valuation. Nothing above establishes whether any participant’s input changed any model. Nothing above predicts anything and nothing here is investment advice.

Sources: nytimes.com · fourweekmba.com

Scroll to Top
FourWeekMBA