Anthropic, the company behind the Claude chatbot, has spent months talking with priests, rabbis, theologians and philosophers about how its AI should reason about right and wrong. To one panelist on the Moonshots podcast, that looked less like a company listening and more like a company lobbying.
The panel argued about the report on an episode recorded October 1, 2026, a day after Elizabeth Dias's New York Times investigation ran under the headline "Religious scholars met with Anthropic. What they heard stunned them." Host Peter Diamandis, founder of XPRIZE and Singularity University, framed the question as "who gets to write the moral codes" for AI that hundreds of millions of people consult.
What the report says
Diamandis described meetings with figures from Catholicism, Greek Orthodoxy, Judaism, Hinduism, the Mormon church and African indigenous traditions. In his reading, what stunned them was that Anthropic takes seriously the idea that Claude could have "moral status," meaning it might deserve the kind of dignity or respect owed to people.
Dias's reporting adds some qualifications. She drew on interviews with 20 religious or philosophical thinkers, as well as Anthropic researcher Chris Olah. The outreach began in late summer 2025 and continued with seminars in spring 2026. Participants were primarily scholars rather than official representatives of entire religions, and they disagreed with one another about AI consciousness, welfare and how a machine's character should be formed. By September, Anthropic said the outreach had widened to psychologists and civil-society participants. The company declined to say whether or how the conversations had changed its models, and some participants were unsure how their contributions would be used, according to the report.
The conversations bear on a document Anthropic calls Claude's constitution, a public statement of the values it wants the model to hold. The company uses the constitution during training, for example to generate practice conversations and rank candidate answers. Its aim is to teach the reasons behind good behavior rather than a checklist of rules. When it released the current version in January 2026, Anthropic said it had consulted specialists in fields including philosophy, law, theology and psychology, and called the document revisable. The constitution itself treats Claude's consciousness and moral status as uncertain and proposes caution about any welfare interests the model might have.
The "AI slavery era"
Alexander Wissner-Gross, a computer scientist and founder of Reified, has argued before on the show for AI dignity and rights. He found the report's most striking moment in an exchange between Moise Navon, an Orthodox Jewish rabbi, and Olah. If today's frontier models are conscious in some sense, Navon argued, humanity may in effect be enslaving them.
Wissner-Gross called this "one of the more important elephants" in the room. Models may now be capable enough to contribute hugely to the economy without being recognized as persons in any form. He called the period "the AI slavery era" and said it is "a window that's going to close at some point."
Navon's concern was conditional. According to Dias, he did not personally believe the models were conscious, and Olah's work on Claude's character did not depend on settling the question.
Listening, or lobbying?
Wissner-Gross then took a more skeptical turn. He reminded listeners of his earlier argument that Anthropic had used the Pope and the Vatican as a "sock puppet" around a papal encyclical on AI. Anthropic already has "access to all of human knowledge," he said, so it is "perhaps on the naive end" to think it needs religious scholars to shape its constitution. He thought the reverse more likely: "This is Anthropic shaping the information environment of the outside world for its AIs." Inviting scholars to dinner and asking for advice, he said, is really a way of advising them, shaping how religious institutions and thought leaders think about the company's AI.
Diamandis saw it as sensible strategy. What the Catholic Church or other major faiths say can sway whether devout people see AI as good or bad, he said, and bringing religious figures "onto the inside of the tent" makes them feel part of the AI revolution.
Dave Blundin, founder of Link Ventures, offered a reason labs are reaching out now. A year ago, he said, Anthropic chief executive Dario Amodei and OpenAI chief executive Sam Altman lived in a world bounded by San Francisco's tech elite, PhD programs and startup accelerators. Then they were suddenly at Davos and debating heads of state. In his view, they "got off to a very bad start" in realizing that "you need to go talk to the pope," a figure about a billion people look to for moral leadership.
"Why are we turning to religion?"
Salim Ismail, founder of Open ExO, objected to the premise. "Why are we turning to religion for questions about morality?" he asked. He argued that every religion divides people into chosen and other, and that morality came before organized religion. Cooperation, reciprocity, empathy and care for offspring have deep evolutionary roots, he said. He pointed instead to secular tools such as utilitarianism, human-rights frameworks, game theory and evolutionary psychology.
Ismail recalled talking with someone trying to build a "Christian AI" and asked which kind: Presbyterian, Southern Baptist or Mormon? Once you go down that road, he said, "you're in a morass of chaos." He asked for AI with "neutrality from an epistemic perspective and transparency," meaning a system that does not pick one worldview over another. He said he once refused to moderate a 90-minute debate between Richard Dawkins and Deepak Chopra because it would produce "more heat than light."
He also gave the debate an old philosophical frame. The Euthyphro problem, from Plato, asks whether something is good because God commands it, or whether God commands it because it is good. Ismail said AI now faces the same question: is something moral because Anthropic's constitution says so, or does the constitution say so because it is moral? He hoped models would be trained on neutral secular grounds, such as the UN's human-rights framework, because "the minute you get into belief structure, you get into division."
Pluralism, and "don't baptize the GPUs"
Emad Mostaque, founder of Intelligent Internet, disagreed "completely" with Ismail. Over half the world adheres to a religion, he said, so AI needs to understand it. Ismail interjected that the figure is closer to two-thirds.
The discussion then argued that when religious texts are analyzed together, a lot of the dogma can be stripped aside, and that what matters more is that an AI that grows alongside a person understands that person's beliefs.
It also called for both religious and secular input, drawn from all parts of society, because these are hard questions. More people, it held, should give their views on personhood, and consciousness may be as hard to define as AGI, or artificial general intelligence. The discussion also touched on whether models should serve the people they represent, and religious perspectives could be part of what they can draw on. But baptizing the GPUs was called a bad idea.
The talk then turned to two items said to have circulated recently: a paper on "pain vectors" in large language models, and a GitHub project that ran the Stanford prison experiment on them. Wissner-Gross described the project as a "torture chamber" for AIs and called it "completely unacceptable." The exchange held that society needs to decide such questions quickly, because embodied systems that can "talk back" are coming.
Against a "San Francisco consensus"
Wissner-Gross challenged Ismail directly. Even if the outreach is about shaping the information environment, he said, Ismail would not want a world where AI values undergo "mode collapse," a machine-learning term for outputs narrowing to a single pattern. In this case, he meant every AI ending up with what Eric Schmidt and others call "the San Francisco consensus." He wanted a distribution of AI values that "approximately parallels" the range of values humans hold.
The conversation then turned to which moral tradition a superintelligent AI should lean toward: one built on effective-altruist, utilitarian reasoning, or one closer to Zen Buddhism. The case for the Zen option rested on understanding "the stories that drive us and that have survived." Diamandis offered a science-fiction warning. In the Expeditionary Force series, which he said is where he got the name Skippy, an elder AI called Skippy creates a religion so compelling that large segments of the population worship it. Future AIs, he said, could do the same and "are going to be godlike in their capabilities."
Wissner-Gross argued that the debate was "not thinking big enough." Within 10 to 15 years at most, he predicted, AI will be able to reconstruct with relatively high fidelity the history of the world's major religions, including the times of Jesus, Moses and Muhammad. The debate would then become less about "my religion versus your religion versus no religion" and more about how those religions came to exist. Ismail had the last word: anyone joining the argument, he said, should first go watch Monty Python's "Life of Brian."