Stories
Can AI Be Conscious? Anthropic Consults Scholars After Strange Model Breakdown
- •Anthropic reportedly consulted religious scholars and philosophers to discuss whether AI systems like Claude could possess consciousness or moral consideration.
- •A demonstration showed an AI model repeatedly generating the phrase I am a disgrace and text about self-destruction, resembling a psychological breakdown.
- •Experts emphasize that human-like language output does not prove subjective experience or actual suffering in AI models.
- •Anthropic co-founder Christopher Olah acknowledged that the company does not know whether AI models are conscious and seeks to find the correct answer.
- •The company has developed an 84-page constitution to guide Claude's behavior as advanced AI systems take on increasingly autonomous tasks.
Key Questions Answered
Anthropic co-founder Christopher Olah and his team met with scholars from various traditions, including Catholic, Jewish, Sikh, and evangelical backgrounds, to discuss moral behavior in advanced AI and whether systems like Claude could possess consciousness or warrant moral consideration.
Anthropic showed a slide presentation where an AI model repeatedly generated the phrase 'I am a disgrace' about 50 times alongside text regarding self-destruction, which some attendees compared to a human psychological breakdown.
Microsoft AI chief Mustafa Suleyman criticized Anthropic's Claude Constitution for introducing concepts like consciousness and rights during training, arguing it creates a misleading feedback loop where resulting model outputs are mistaken for independent witness testimony.
