---
type: "Evidence Item"
title: "From hard refusals to safe-completions: toward output-centric safety training"
description: "Discover how OpenAI's new safe-completions approach in GPT-5 improves both safety and helpfulness in AI responses—moving beyond hard refusals to nuanced, output-centric safety."
resource: "https://openai.com/index/gpt-5-safe-completions"
tags: ["appendix-iii", "vendor", "openai"]
timestamp: "2025-08-07"
category: "vendor"
publisher: "OpenAI"
cope_score: 48
confidence: 0.9
---

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

# Claim

Discover how OpenAI's new safe-completions approach in GPT-5 improves both safety and helpfulness in AI responses—moving beyond hard refusals to nuanced, output-centric safety training for handling dual-use prompts.

# Relevance

Appendix III, section two: vendor threshold and platform capability evidence

# Oracle Verdict

This is a low-signal vendor radar item. Keep it as context only unless a later benchmark, deployment, procurement change, or labour-market datapoint turns it into direct Appendix III evidence.

# Metadata

* Publisher: OpenAI
* Category: vendor
* Sector: AI infrastructure
* Capability: Frontier model release and benchmark movement
* Cope score: 48
* Confidence: 0.9

# Related Concepts

* [Live evidence index](index.md)
* [Thesis](../thesis.md)

# Citations

[1] [From hard refusals to safe-completions: toward output-centric safety training](https://openai.com/index/gpt-5-safe-completions)