From hard refusals to safe-completions: toward output-centric safety training
Discover how OpenAI's new safe-completions approach in GPT-5 improves both safety and helpfulness in AI responsesβmoving beyond hard refusals to nuanced, output-centric safety training for handling dual-use prompts.
Log in to bookmark articles and create collections
Isabella News