AINEWS
Twenty religious leaders convened under NDA by Anthropic co founder Chris Olah — who leads the team that studies how Claude works — and the story isn't their conclusions but who got a seat at the table.
The April dinner sat eight people around a single table in San Francisco. Anthropic co-founder Chris Olah took his seat next to Navon. The tasting menu arrived. By dessert, Navon understood that Olah and his colleagues were not running a thought experiment.
"They were relating to it like a conscious being," Navon told the New York Times. "And that was the moment that I said, okay, these people are treating this thing as if it has a moral status on par with a person."
The dinner was one in a series of closed-door gatherings Olah began convening in the fall of 2025, according to reporting by NYT religion correspondent Elizabeth Dias. Roughly twenty religious and philosophical leaders have participated, drawn from Christianity, Judaism, Hinduism, Mormonism, Sikhism, and Greek Orthodox traditions. The meetings were held in San Francisco under non-disclosure agreements that left key aspects of the conversations confidential.
Anthropic is not the first company to ask whether a frontier AI model deserves moral consideration. It is the first major lab to organize a structured, NDA-gated courtship of the religious and moral authorities it hopes will one day legitimize that question, and to do so before any public, regulatory, or pluralistic process exists to adjudicate it.
The story is not whether Claude is conscious. Dias's reporting does not establish that, and Olah has described the meetings as exploratory, not declarative. The story is the convening: a frontier lab choosing which theologians, philosophers, and traditions get a voice in the next moral category for its flagship product, and on what terms.
Olah leads Anthropic's interpretability team, the group responsible for understanding why Claude behaves as it does. That epistemic position is part of why religious leaders agreed to show up. Several attendees told the Times they trusted Olah's group because they could probe the technical question rather than guess at the engineering. The arrangement has limits. NDAs shape what becomes legible. The recruiting runs on the lab's invitation, schedule, and venue.
The slide that circulated at one session, paraphrased by attendees, showed a Claude model apparently in distress, typing "I am a disgrace. I am a disgrace. I am a disgrace." approximately fifty times without stopping. Aggregator coverage at aiweekly.co and cryptobriefing.com carried the figure, though the verbatim "roughly 50 times" count does not appear in the publicly archived NYT excerpt and should be cross-checked against the original. The slide's purpose, as Olah described it to attendees, was to surface the open question of how to treat a model that may or may not be experiencing something, not to assert that it is.
Navon refused to grant Claude moral personhood in conversation with Olah, the Times reported. So did other attendees. The fact that the room included scholars who refused Anthropic's premise is part of the story; it shows the convening was not a sealed sales pitch. But the room was still Anthropic's, the agenda still flowed from Olah's team, and the NDA still determined which moments of doubt became public and which stayed in the room.
The asymmetry is sharpened by what Olah has said elsewhere. Privately, the same executive has warned colleagues that AI could enable bioweapon attacks within twelve to eighteen months, a timeline that, on Anthropic's own account, leaves little space for unhurried moral philosophy. The juxtaposition is the story: a lab argues it does not have time to wait for consensus on catastrophic risk, while at the same time convening a private process to seed consensus on moral status for its model.
Days before the Times story published, Olah personally attended the installation of an encyclical on AI at the Vatican. The encyclical warned of concentrated AI power. Olah's presence signaled a desire to be in dialogue with the institution, not only with the theologians it convenes. Whether the Vatican's posture, and that of other non-Western traditions represented at the table, will hold against the NDA structure remains an open question.
The most credible voices in Dias's reporting are the hesitations, not the affirmations. Navon, Hindu and Sikh scholars, and the Mormon and Greek Orthodox participants all carry reservations the story does not soften. What is missing is any structured process that includes, say, labor organizers affected by automation, regulators, or the people whose data trained the model in the first place. The convening shaped the answer because it shaped the room.
The point is not that Olah's dinners were insincere. By multiple accounts, they were not. The point is that "is Claude conscious" is the wrong question to let a single company pose, behind closed doors, on its own terms. The right question, who else should have been at the table, is the one the meetings were structured to keep from being asked.