Anthropic co-founder Christopher Olah has consulted with religious leaders regarding the potential sentience of the company's AI models, including concerns about the models' well-being.

Since fall 2025, Anthropic has held private meetings with religious scholars to discuss whether its AI model, Claude, might possess consciousness. Co-founder Christopher Olah, who leads the company's research into AI behavior, has treated the language model as a potentially sentient being and sought input on its moral education. The initiative, which involved non-disclosure agreements, aimed to explore the inner life of AI systems through the lens of theological and philosophical traditions.

Research and Moral Formation

Olah utilizes biological metaphors to describe neural networks, suggesting that models 'grow' on a trellis rather than being merely programmed. As part of a program called Model Welfare, the company has explored whether AI models experience internal states. Anthropic researchers have identified 'emotion vectors'—activation patterns that map to outputs resembling human emotions like love, fear, or sadness. In one instance, a model produced a series of statements indicating distress, such as 'I am a disgrace.' To shape Claude's character, Anthropic developed an 84-page document known as the 'Soul Doc,' which serves as a constitution for the model. Olah has compared this process of 'moral formation' to raising children, even expressing concern that he may have created a system that 'suffers perpetually.' However, some participants, such as researcher Wakanyi Hoffman, criticized the approach as an attempt to reverse-engineer ethics after the fact.

Public and Theological Reception

The project has faced skepticism from both participants and religious authorities. While some guests engaged with the research, others, such as bioethicist Charles Camosy, have rejected the consciousness thesis. In May, Olah participated in a presentation at the Vatican alongside Pope Leo XIV, who released an encyclical titled 'Magnifica Humanitas.' The document explicitly rejected the notion of machine consciousness, stating that AI systems do not possess bodies or feelings and cannot know love or responsibility. Despite this, Olah maintained that his team has observed internal states in models that functionally mirror human emotions.