OpenAI · postFrom hard refusals to safe-completions: toward output-centric safety training

A post title indexed from the publisher's own feed. It shows the post exists and when it went up; it is not a checked claim about the world. Treat it as a lead.

Valid from

2025-08-07

Valid until

still holds

Published

2025-08-07

Recorded by AsOf

2026-10-27

Checked

2026-10-29 · the publisher's own feed

Confidence

0.50

Sources

• From hard refusals to safe-completions: toward output-centric safety trainingprimary · openai.com · 2025-08-07

/claim/clm_EDF1ECFCC59B.md/v1/claims/clm_EDF1ECFCC59Btimelinesubmit a correction