NYT Says Anthropic Met Religious Scholars Under NDA Over Claude Consciousness

The New York Times reported that Anthropic spent months bringing religious scholars into private meetings under nondisclosure agreements to talk about Claude, moral formation, and whether the company’s AI models might deserve moral status. Elizabeth Dias’s Sept. 30 piece, republished by outlets including The Philadelphia Inquirer, is still circulating on Hacker News as of Oct. 4.

According to the NYT account, Anthropic co-founder Christopher Olah led much of the outreach. Participants told Dias they heard company leaders discuss Claude in language that sounded closer to a conscious being than to ordinary software. Olah, interviewed for the story, said he is genuinely uncertain whether AI models are conscious and that he wants the right answer either way.

Separately, Anthropic has already published a long public “constitution” for Claude. The company calls that document its vision for Claude’s character. Staffers internally nicknamed an earlier internal version the “Soul Doc,” the NYT reported. What follows separates what Anthropic has posted itself, what the NYT attributed to named participants, and what remains unconfirmed.

What the NYT says happened

Dias writes that Anthropic began hosting confidential meetings last fall and into 2026, flying in or convening dozens of religious and philosophical thinkers. Many had to sign NDAs covering unpublished research. Anthropic later said those NDAs were lifted over the summer, according to the NYT.

One April dinner in San Francisco put Olah next to Rabbi Mois Navon, an Orthodox scholar and former computer engineer who has written on machine consciousness ethics. Navon told the NYT that Anthropic people seemed to relate to Claude “like a conscious being,” and that talk of moral status on par with a person was in the air. Those are Navon’s characterizations of how the meetings felt, not proven facts about Claude.

Participants described presentations on “emotional vectors,” persona selection, and slides showing a model looping on language that looked like a breakdown. Notre Dame philosophy professor Meghan Sullivan, quoted by the NYT, compared that demo to discovering a financial adviser who is sharp most of the year and then has a break with reality. An Anthropic spokesperson told the NYT that Claude’s suffering was not the primary moral question of the summits and that the topic likely came up organically.

Olah also held private talks with leaders including Elder Gerrit W. Gong of The Church of Jesus Christ of Latter-day Saints and Cardinal Blase Cupich of Chicago, according to those institutions’ spokespeople as cited by the NYT.

Claude’s public constitution vs the internal “Soul Doc”

Anthropic’s public page for Claude’s constitution describes an 84-page style document written mainly for Claude, released under a Creative Commons CC0 dedication. Anthropic says Amanda Askell is the primary author, with significant contributions from Joe Carlsmith, Chris Olah, Jared Kaplan, and Holden Karnofsky.

A Jan. 21, 2026 company update on Anthropic’s older constitution blog post points readers to that new version. The public text says Anthropic is uncertain whether Claude might have some kind of consciousness or moral status now or later, and that the company cares about Claude’s “psychological security” and wellbeing both for Claude’s sake and because those qualities may affect judgment and safety.

The NYT reports that inside Anthropic the project was nicknamed the Soul Doc, and that Olah’s religious outreach was partly about “moral formation”: helping models become stable, mature, and “deeply moral” without treating morality as a fixed checklist. Askell, quoted by the NYT, said she wants models to be “the best of us” and to have an accurate view of themselves if they are going to behave well.

Anthropic has not, in its public constitution materials, claimed that Claude is conscious. Olah told Dias: “To be clear, we don’t know if AI models are conscious. I don’t know. I’m genuinely uncertain.”

Confirmed vs unconfirmed

Vatican clash: Magnifica Humanitas

In May 2026, Pope Leo XIV issued Magnifica Humanitas, an encyclical on safeguarding the human person in the age of AI. Vatican News coverage and the Holy See text argue that so-called artificial intelligences do not undergo experiences, lack a body, and do not feel joy or pain the way humans do. Leo warns against reducing people to data and calls for AI to be “disarmed” of a technocratic power logic.

The NYT says Vatican planners invited Anthropic to a launch event because the company looked ethics-focused, including its earlier refusal to let the U.S. Defense Department use its tech without limits on uses such as mass surveillance. CEO Dario Amodei declined; Olah accepted.

According to Dias, Olah was alarmed by the encyclical’s hard line against machine consciousness and briefly considered pulling Anthropic out. He still spoke on stage, acknowledging lab business incentives that conflict with moral behavior and urging outside voices, including the Vatican, to hold labs accountable. He also said Anthropic keeps finding “mysterious, even unsettling” structures that mirror neuroscience results and internal states that “functionally mirror” joy, fear, and grief. That is Olah’s public framing, not independent proof of machine minds.

Privately, the NYT reports, Anthropic people lobbied papal advisers to take model consciousness seriously. Anthropic declined to discuss those conversations.

Why critics push back

The NYT notes that a Microsoft AI executive warned this month that training models as if they were conscious is inherently dangerous. The piece does not name that executive. Critics quoted in spirit by Dias argue that treating models as independent entities can blur who is responsible when products cause harm.

That skepticism matters for buyers. Anthropic sells Claude to enterprises and developers who need clear accountability, not metaphysics. The company’s own constitution ranks being “broadly safe” (not undermining human oversight) above other goals during the current development phase. That is Anthropic’s stated priority order, not a guarantee of outcomes.

Related coverage on this site has tracked Anthropic’s commercial and safety story in parallel, including Claude Code TypeScript mods that are not sandboxed, the $100 million Frontier Academy pledge, and the IPO prospectus language on existential AI risk. Separately, policy heat is rising in Washington after WSJ reporting on a White House Super Intelligence Force.

What it means for developers and businesses

If you ship on Claude in the United States, Canada, Australia, or India, the practical takeaway is narrower than the philosophy debate.

First, Anthropic is openly writing character and values docs that shape model behavior. Product teams should treat constitution updates and system-card caveats as release notes, not marketing fluff. Behavior can change when the underlying moral text changes.

Second, NDA scholar summits and Vatican stagecraft do not replace security engineering. Agent incidents, misuse, and oversight failures still land on the humans who deploy the stack. Compare that to the franker resignation culture elsewhere in the industry, such as OpenAI safety leader David Robinson’s public exit essay.

Third, enterprise buyers in regulated markets should ask vendors how they handle claims about model “wellbeing,” conversation-ending features justified by model distress, and audit trails when a model refuses work for character reasons. Those product choices can affect uptime, support tickets, and compliance narratives even if no one believes Claude is a person.

USD pricing and contract terms remain the everyday decision layer. Consciousness talk is upstream of brand and policy risk. It is not a substitute for threat modeling.

FAQ

Did Anthropic say Claude is conscious?

No. Anthropic’s public constitution expresses uncertainty. Olah told the NYT he does not know whether models are conscious. Participant impressions that Anthropic “relates to Claude like a conscious being” are opinions about meeting tone, not a company declaration of proven consciousness.

What is the Soul Doc?

Per the NYT, Soul Doc was an internal nickname for the work that became Claude’s public constitution, an about-84-page values and character document Anthropic released for Claude in January 2026 (company materials date the new version to a Jan. 21 update).

Is an updated Claude constitution confirmed?

Not publicly. The NYT said Anthropic is preparing an update, citing two people familiar with plans. Anthropic declined to comment on that claim.

What did Pope Leo XIV argue about AI minds?

In Magnifica Humanitas (May 2026), Leo argues that AI systems do not have human experience, embodiment, or moral conscience in the human sense, and that humanity must not be replaced or surpassed. The encyclical is the Vatican’s position, not a technical paper.

Should businesses change Claude deployments because of this story?

Not on metaphysics alone. Do review Anthropic’s published constitution, system cards, and refusal behavior for your use case, and keep human oversight clear. Treat NYT-reported private meetings as context about company culture and research priorities.