Safe-completions approach in GPT-5: moving beyond hard refusals to output-centric safety training (openai.com)

OpenAI research on a safety training method that generates nuanced, helpful responses to dual-use prompts instead of refusing them outright, balancing safety guardrails with model usefulness.