From hard refusals to safe-completions: toward output-centric safety training

Discover how OpenAI’s new safe-completions strategy in GPT-5 improves each safety and helpfulness in AI responses—shifting past hard refusals to nuanced, output-centric safety training for dealing with dual-use prompts.