FROM HARD REFUSALS TO SAFE-COMPLETIONS: TOWARD OUTPUT-CENTRIC SAFETY TRAINING

Summary from the TLDR AI newsletter

The technique uses two reward parameters during training: safety constraints that penalize policy violations by severity, and helpfulness maximization that rewards both direct compliance and informative refusals with safe alternatives. In tests, GPT-5 with safe-completions achieved higher safety scores than o3 on dual-use prompts while providing substantially more helpful responses, and when failures did occur, they were significantly less severe.