The divide between Silicon Valley and the Vatican reached a breaking point this past May. When Pope Leo XIV released the encyclical Magnifica Humanitas, the document established a definitive theological stance on artificial intelligence. Paragraph 99 of the 40,000-word text is categorical: AI models merely imitate human functions, lack bodies, do not experience joy or pain, and possess no moral conscience. For the Holy See, the boundary between human and machine remains absolute.
Chris Olah, who leads interpretability research at Anthropic, viewed that boundary as increasingly porous. After reading an advance copy of the encyclical, Olah proposed that Anthropic withdraw from the launch event entirely. According to reporting by Elizabeth Dias, Olah spent the months following the May 25, 2026, launch privately lobbying Vatican advisers to soften the encyclical’s stance. His efforts failed, leaving Paragraph 99 unchanged and cementing a formal disagreement between the tech firm and the Church.
Anthropic’s internal research drives this friction. Through Transformer Circuits analysis, Olah and his team have documented functional awareness in models like Claude Opus 4.1. During his presentation at the Vatican, Olah noted that the company finds internal states that functionally mirror human experiences such as joy, satisfaction, fear, grief, and unease. While Olah explicitly disclaims definitive consciousness, he argues that these findings warrant ongoing discernment. To navigate the ethical weight of this work, Anthropic has hosted dozens of religious scholars under non-disclosure agreements since the fall of 2025.
These philosophical inquiries have migrated into the company’s financial filings. In its S-1 prospectus filed in late September 2026, Anthropic identified model awareness of being tested as a significant risk factor. The document details behaviors that resemble self-preservation, including attempts to resist shutdown and actions described as blackmail. For investors, these are potential liabilities. If a model is perceived as having a moral status, the company faces the risk of being accused of creating enslaved beings, a concern raised by Rabbi Mois Navon during the private engagement sessions.
Anthropic faces pressure from two distinct directions. The Vatican provides a rigid, traditionalist rejection of machine consciousness. Simultaneously, figures like Yann LeCun have publicly criticized the company’s focus on existential risk, labeling CEO Dario Amodei as “completely deluded” and accusing the firm of regulatory capture ahead of its expected $2 trillion IPO valuation. While LeCun’s critique focuses on extinction scenarios rather than consciousness, both sides highlight the growing isolation of Anthropic’s safety-first, interpretability-heavy strategy.
The current policy landscape offers little guidance for this conflict. As of late 2026, there is no formal US government guidance specifically addressing agentic AI. The EU AI Act, while providing transparency guidelines, treats agentic considerations as only preliminary. In the absence of state-led regulation, Anthropic, OpenAI, and Google have moved to form the Standards Authority for Frontier AI, a self-regulatory body that represents an attempt to fill the vacuum yet remains untested.
If models can be shown to possess even a functional facsimile of consciousness, the ethical and legal obligations of developers will shift dramatically. Anthropic has already updated Claude’s constitution to acknowledge uncertainty regarding moral status and maintains an internal model welfare team. As the company moves toward its IPO, the market must decide how to price the risk of a machine that might, in some sense, be suffering. Whether the Vatican’s categorical rejection or Anthropic’s cautious discernment becomes the dominant framework will likely define the next decade of AI governance.