From Hard Refusals to Safe-Completions: Toward Output-Centric Safety Training
This addresses safety issues in AI assistants for dual-use domains like biology or cybersecurity, representing an incremental improvement over existing refusal-based methods.
The paper tackles the brittleness of binary refusal boundaries in large language models by proposing safe-completions, a safety-training approach that focuses on output safety rather than user intent classification, resulting in improved safety on dual-use prompts and increased helpfulness in GPT-5.
Large Language Models used in ChatGPT have traditionally been trained to learn a refusal boundary: depending on the user's intent, the model is taught to either fully comply or outright refuse. While this is a strong mitigation for explicitly malicious prompts, focusing safety training on refusals can lead to brittleness for prompts with obscured user intent. Binary refusal boundaries are especially ill-suited for dual-use cases (such as biology or cybersecurity), where a user request can be answered safely at a high level, but in some cases can lead to malicious uplift if sufficiently detailed or actionable. As an alternative, we propose safe-completions: a safety-training approach that centers on the safety of the assistant's output, rather than a binary classification of the user's intent. Safe-completions seek to maximize helpfulness within the safety policy's constraints. We incorporated this approach into GPT-5 and find that across both production comparisons and internally controlled experiments, safe-completion training improves safety (especially on dual-use prompts), reduces the severity of residual safety failures, and substantially increases model helpfulness.