OpenAI00:00PolicyOfficial Blog
OpenAI Reveals Goblin Cause & Adds Codex Suppression
Self-suppress quirks for stable business code.
Key Points
- 1'Nerdy' rewards overfit goblins
- 2Future: remove signals, filter data
- 3Codex suppression prompts added
- 4Audit techniques for prod use
Goblin mentions in GPT-5.1 stemmed from training reward bias in 'Nerdy' personality. Fixed by removing signals and filtering data for future models. Codex now has dev prompts to suppress. Teaches auditing model quirks.