Anthropic co-founder Christopher Olah met with Pope Leo XIV at the Vatican this month to discuss the dangers of advanced artificial intelligence [1, 2].
The meeting signals an unusual alliance between a leading AI laboratory and the Catholic Church to address the existential and ethical risks of machine learning. As AI models grow more complex, the push for a global framework to constrain the technology is gaining momentum from both technical and spiritual leaders.
The discussions followed the release of a new AI encyclical by Pope Leo XIV [3, 5]. During the meeting, Olah raised concerns regarding the internal workings of large-scale models, describing certain observed phenomena as disturbing [1, 2].
"We’re finding unsettling things inside AI models," Olah said [2].
Olah said the internal processes of these systems are mysterious, noting that what happens inside an AI's mind is unsettling [4]. These observations align with the Pope's own concerns regarding the trajectory of the technology. According to reports, the Pope called for "disarming" the technology [1].
The collaboration between Anthropic and the Vatican suggests a shared goal of increasing transparency, and safety measures for AI development [3]. By investigating the "mysterious" nature of AI, Olah and the Pope aim to identify specific risks before they manifest in public-facing applications [4].
This dialogue occurs as the Church seeks to provide moral guidance on the use of automation and intelligence. The focus remains on the need to prevent AI from becoming a tool of destabilization, or dehumanization [5].
“"We’re finding unsettling things inside AI models."”
This meeting represents a convergence of technical interpretability research and global moral authority. By framing AI safety as a need to 'disarm,' the Vatican and Anthropic are treating advanced AI not merely as a software challenge, but as a systemic risk akin to weaponry, suggesting that voluntary corporate safety guidelines may be insufficient without a broader ethical or regulatory mandate.
