From outright refusals to safe completions: advancing output-focused safety training
OpenAI's new safe completions method for GPT-5 replaces the traditional hard refusal approach with output-centric safety training. Instead of refusing potentially sensitive questions outright, the mod…