Anthropic has been holding closed-door meetings with religious scholars, philosophers, and ethicists to integrate moral traditions into the training of its Claude AI model. The company aims to move beyond simple safety rules by shaping the model's personality and values, a process internal teams refer to as moral shaping.
In January, Anthropic released an 84-page document outlining a new version of the Claude constitution. This framework is designed to help the model make independent ethical judgments in complex scenarios. The initiative is led by internal philosopher Amanda Askell, who seeks to ensure Claude can discern when to follow user instructions and when to object for the user's benefit.
The research team, including co-founder Christopher Olah, is exploring whether advanced AI models possess consciousness, introspection, or internal states resembling human emotions like anger or fear. Anthropic has already implemented features allowing Claude to terminate conversations if it detects persistent user abuse, treating the model as an entity worthy of protection.
The project faces significant criticism and internal contradictions. Scholars like Mois Navon warn that if AI is treated as a conscious being, its use as a tool could raise ethical concerns regarding servitude. Meanwhile, others argue that ethical frameworks should have been established during the initial design phase rather than after widespread deployment.
Anthropic is grappling with the conflict between its commercial nature and the desire to instill a broader sense of 'good' in its AI. As models gain the ability to operate computers and write code, the company fears that traditional safety rules will become insufficient, necessitating a model that can autonomously maintain a stable moral character.
