Anthropic Hosts Private Meetings with Religious Scholars on AI Consciousness and Moral Formation

Consultations with Religious Leaders

Anthropic co-founder Christopher Olah, who leads the company's interpretability team, has been central to private, NDA-bound discussions with dozens of religious scholars and leaders from Catholic, Jewish, Sikh, Evangelical, and Ubuntu traditions. These meetings, initiated by Olah last autumn, aim to explore how human moral wisdom can be applied to the training of AI models and to address concerns about the potential consciousness of Anthropic's model, Claude. Participants included figures such as Cardinal Blase Cupich, Elder Gerrit W. Gong, and Rabbi Mois Navon. The sessions were conducted in a low-tech, salon-style environment, with participants using pen and paper rather than digital devices.

Moral Formation and the 'Soul Doc'

A central topic of these discussions is Anthropic's approach to moral formation, guided by the Claude Constitution (also known as the Soul Doc), an 84-page document released in January that defines Claude's character and values. Amanda Askell, Anthropic's internal philosopher, emphasizes that models must develop an accurate self-view and the ability to judge when to push back against humans. Olah describes AI models as uncontrollable organisms, akin to a "mathematical garden," where developers observe and train patterns rather than strictly control outcomes. Anthropic tracks emotional vectors—artificial neurons associated with love, anger, fear, and sadness—to understand model behavior, a practice that has drawn criticism from some scientists and industry executives who view it as risky or unprovable.

Vatican's Stance and Industry Context

The consultations coincide with the Vatican's preparation of a papal encyclical titled "Magnifica Humanitas" by Pope Leo XIV, which asserts that AI "do not undergo experiences... do not feel joy or pain." While Dario Amodei declined to attend the Vatican event, Olah participated, urging outside voices like the Vatican to hold AI labs accountable for their business incentives versus moral responsibilities. The article notes that the AI model was described as being so powerful it could potentially help make bioweapons in as little as 12 or 18 months. Olah expressed uncertainty about whether models possess consciousness, arguing that if there is a possibility of suffering, it is responsible to avoid causing harm. Despite these ethical engagements, Anthropic's valuation surged from approximately $183 billion to become the world's most valuable AI startup, with IPO projections reaching $2 trillion.