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

In three linesOpenAI introduces output-centric safety training for GPT-5, replacing hard refusals with nuanced responses. The "safe-completions" approach improves both safety and helpfulness on dual-use prompts, though specific benchmarks and technical details are not disclosed in the excerpt.

Summary generated by Claude — human-verified