Anthropic releases 80-page "Claude Constitution" and upgrades its AI alignment framework.

PANews reported on January 23 that Anthropic, an AI large-scale model company, recently released a new version of its "Claude Constitution" under a Creative Commons CC0 1.0 license. This new constitution serves as the "supreme authority" for training, used to generate synthetic training data and ranking feedback. Instead of simply listing principles, it explains "why this is so," aiming to improve generalization to new scenarios. The document is ordered by broad safety > broad ethics > adherence to guidelines > genuine assistance, listing "hard constraints" (such as not providing substantial assistance in biological weapons development) and adding chapters on virtues, psychological safety, and model self-awareness, emphasizing transparency and continuous iteration.