FrontierThe story, in brief
From hard refusals to safe-completions: toward output-centric safety training
GPT-5 ditches hard refusals. OpenAI's new 'safe-completions' approach trains models to answer risky questions—safely.
The KeyNews take
Why it matters
OpenAI is redefining AI safety training away from blocking dangerous requests toward nuanced output control. This shifts how enterprises balance security with usability, and sets a new standard for handling dual-use prompts that competitors will need to match.
The key facts
5 to knowGPT-5 introduces 'safe-completions' safety training methodology
Moves from hard refusals to output-centric safety approach
Improves both safety and helpfulness metrics
Designed to handle dual-use prompts with nuance
Published August 7, 2025 on OpenAI official channel
Go to the source
OpenAI Blogopenai.com
Publisher excerpt: 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.