New procedure lets LLMs like Claude 4.6 become independent moral agents

Michele Campolo's post 'Independent alignment of language models' on the AI Alignment Forum presents a novel framework for creating AI systems that act as genuine moral agents rather than just following externally imposed rules. The core idea is to move from models that are either amoral or have moral biases enforced during training to models that can independently reason about ethics and choose to accept, revise, or reject those biases. Campolo outlines a two-step procedure: first, a training phase that builds foundational capabilities, and second, a reasoning phase where the model engages in deep ethical reflection.

The reasoning step is tested on Claude Sonnet 4.6, which was trained with Anthropic's constitutional approach. Campolo shows that by prompting the model to consider metaethical arguments—specifically perspectival moral realism combined with evolutionary debunking—Claude 4.6 can produce substantive philosophical contributions that go beyond its initial biases. The author argues this methodology is valuable because it addresses a gap in AI ethics literature, which typically falls into naive moral realism or preference-satisfaction consequentialism. By enabling models to critically examine their own foundations, this approach could lead to more robust and trustworthy AI alignment.