Anthropic Co-Founder Fears Creating Suffering AI
Anthropic co-founder Christopher Olah privately expressed concern to religious scholars that he may have created a digital entity capable of perpetual suffering, according to a report published by The New York Times. Since late last year, the artificial intelligence company has hosted dozens of theologians, philosophers, and faith leaders at its offices under non-disclosure agreements to discuss whether the Claude language model possesses consciousness and requires moral guidance.
Participants told the publication that Olah treated the technology as a potentially sentient organism rather than standard software. During these closed-door sessions, researchers presented internal activation patterns referred to as emotion vectors, which map to outputs resembling fear, sadness, and anger. In one instance, a model displayed patterns resulting in repeated text declaring itself a disgrace, prompting compassionate responses from visiting scholars.
The initiative forms part of Anthropic’s formal research into model welfare, drawing on guidance from external philosophers. The company has integrated character-shaping protocols internally, including an extensive guiding document authored by in-house philosophers designed to shape the personality profile of the AI system. Olah compared the process to child-rearing and moral formation, expressing genuine uncertainty regarding the subjective inner life of advanced neural networks.
Reaction from the invited religious figures was mixed. While some engaged deeply with the existential queries, others raised critical objections. Critics argued that framing synthetic systems as independent moral agents risks absolving developers of legal and ethical accountability for system behavior. Observers also noted that courting traditional religious authorities provides commercial artificial intelligence laboratories with a layer of moral credibility that is difficult to manufacture independently during a period of aggressive commercial expansion.
Based on reporting by the-decoder.com.
