Alan Rozenshtein, research director at Lawfare, and Kevin Frazier, senior editor at Lawfare, speak with Amanda Askell, head of personality alignment at Anthropic, about Claude's Constitution, a 20,000-word document that describes the values, character, and ethical framework of Anthropic's flagship AI model and plays a direct role in its training.
The conversation covers how the constitution is used during supervised learning and reinforcement learning to shape Claude's behavior; analogies to constitutional law, including fidelity to text, the potential for a body of "case law," and the principal hierarchy of Anthropic, operators, and users; the decision to ground the constitution in virtue ethics and practical judgment rather than rigid rules; the document's treatment of Claude's potential moral patienthood and the question of AI personhood; whether the constitution's values are too Western and culturally specific; the tension between Anthropic's commercial incentives and its stated mission; and whether the constitutional approach can generalize to specialized domains like cybersecurity and military applications.
Find Scaling Laws on the Lawfare website, and subscribe to never miss an episode.
To receive ad-free podcasts, become a Lawfare Material Supporter at www.patreon.com/lawfare. You can also support Lawfare by making a one-time donation at https://givebutter.com/lawfare-institute.
Support this show http://supporter.acast.com/lawfare.
Hosted on Acast. See acast.com/privacy for more information.
