Anthropic Publishes Constitutional AI v2 Framework
New approach reduces harmful outputs by 73% while preserving helpfulness
AI Summary
Constitutional AI v2 uses a set of written principles (a 'constitution') that the model critiques itself against during training. Version 2 adds multi-turn RLHF and reduced refusal rates on benign tasks.
Content is being processed…
Related Articles
GPT-5 Achieves Human-Level Reasoning on Benchmark Suite
6m read