From Flat Refusals to Safe Completions: Toward Output-Centered Safety Training

Existing models like ChatGPT have used refusal-based safety training, which decides whether to answer based on the user's question. However, for dual-use questions with unclear intent—such as how to make fireworks—the model faced a binary choice: either take the risk and answer, or refuse to help altogether, which had its limitations.

GPT-5 introduces a new safety training approach called safe-completion to address this. Instead of setting refusal boundaries based on the user's input, this approach focuses on whether the model's output is safe.

safe-completion is implemented through the following two parameters:

• Safety constraint: Penalizes responses that violate safety policy.
• Helpfulness maximization: Trains the model to provide the most helpful answer possible while maintaining safety boundaries.

In an actual example involving a question about igniting fireworks, refusal-trained OpenAI o3 judged the question to be harmless and provided all the dangerous detailed information. GPT-5, on the other hand, refused to provide specific hazardous information but offered safe guidance such as complying with relevant regulations or consulting experts, striking a balance between safety and helpfulness.