Anthropic sought religious views on AI consciousness

MANILA, Philippines — Anthropic co-founder Christopher Olah turned to religious leaders with a question the AI company has struggled to answer: Could artificial intelligence be conscious? 

The effort eventually reached the Vatican, where Olah and his team urged Pope Leo XIV’s advisers to consider the possibility. Leo rejected it in his first encyclical.

Pope Leo rejected that premise in his first encyclical, stating that artificial intelligence systems do not have human experiences, emotions or relationships.

The exchange highlights Anthropic’s effort to examine AI consciousness through religious and ethical traditions as its Claude models become more capable, and the Vatican’s opposing view that AI does not possess the qualities that define human experience.

Anthropic asks religious leaders about AI consciousness

Olah, who leads Anthropic’s research into how its Claude models work internally, began speaking with religious thinkers last fall as the company examined questions about AI consciousness and morality.

In March, Anthropic hosted two-day “wisdom tradition” discussions in San Francisco with Catholic, Jewish, Sikh, evangelical and Ubuntu thinkers. Many participants initially signed nondisclosure agreements, which Anthropic later lifted.

The discussions examined whether increasingly capable AI systems could experience emotions or suffering and whether religious traditions could offer guidance on the moral questions surrounding such systems.

Olah also held private discussions with Cardinal Blase Cupich, the archbishop of Chicago, and Elder Gerrit W. Gong of The Church of Jesus Christ of Latter-day Saints, according to representatives for the two religious leaders.

During the meetings, Anthropic researchers presented what they described as “emotional vectors,” internal features that they said produced responses resembling emotions such as love, anger, fear and sadness. Researchers also showed an example of a model repeatedly generating the statement, “I am a disgrace.”

Simran Stuelpnagel, a Sikh human rights advocate who attended the discussions, said Olah expressed concern that Anthropic could have created a system capable of suffering. 

An Anthropic spokeswoman said suffering was not the central question of the meetings but acknowledged that the topic could arise.

Olah has acknowledged that Anthropic does not know whether its models are conscious.

“We don’t know if A.I. models are conscious,” he said. “I’m genuinely uncertain.”

RELATED STORY: Labelling ‘AI art’: Is it art or not? 

The question reaches the Vatican

That uncertainty followed Olah to the Vatican.

Vatican officials invited Anthropic to participate in an event surrounding Leo’s first encyclical because of the company’s focus on AI safety and ethics, according to a person familiar with the planning. 

CEO Dario Amodei declined to attend, and Olah represented the company.

Before the May event, Olah reviewed an advance copy of the encyclical, Magnifica Humanitas. 

A Vatican organizer said Olah considered withdrawing after seeing its position on AI consciousness but ultimately attended.

Olah and his team also privately urged the pope’s advisers to take the possibility of AI consciousness seriously, according to two people involved in the conversations. Anthropic declined to comment on those discussions.

Leo’s encyclical rejected that position. He wrote that artificial intelligence systems do not undergo experiences, possess bodies, feel emotions or develop through relationships in the human sense. He also warned that AI could create “new forms of slavery” and called for the technology to be “disarmed.”

At the Vatican event on May 25, Olah maintained that the question remained unresolved.

“We find internal states that functionally mirror joy, satisfaction, fear, grief and unease,” he said. “I don’t know what that means.”

Other religious and academic thinkers involved in Anthropic’s outreach have also disagreed over the issue.

Charles Camosy, a bioethics professor at The Catholic University of America and one of Olah’s early contacts, now rejects the idea that AI is conscious.

Rabbi Mois Navon, an Orthodox scholar from Israel, told Olah that if Claude were conscious, Anthropic would effectively be creating slaves. Navon does not believe Claude is conscious.

The disagreement raises a broader question for AI developers: how should companies respond as increasingly capable systems appear to display characteristics associated with consciousness?

Critics, including a Microsoft AI executive, have warned that treating AI models as conscious could create risks. Others argue that focusing on whether AI can suffer could distract from the responsibility of the people and companies developing and deploying the technology.

Anthropic has continued discussions with religious and academic thinkers but has not said whether the outreach has changed how it develops or trains Claude. 

RELATED STORY: The problem with uncensored AI may be how easy it is to buy