ChatGPT can't stop talking about Goblins
OpenAI published a postmortem on why ChatGPT 5.x keeps describing bugs as "perf gremlins" and devs as "fitness goblins." It's an RL contamination story: the "Nerdy" personality reward leaked into default behavior across training.
The mitigation
- Codex's system prompt literally contains "never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons" — repeated mid-prompt because the model ignored it the first time.
- Frogs were investigated and cleared: usage is "legitimate."
The cause
- The "Nerdy" personality system prompt rewarded creature metaphors; in audits the Nerdy reward scored goblin/gremlin outputs higher in 76.2% of datasets.
- RL doesn't keep a learned tic scoped to its trigger. Once goblins were preferred, supervised fine-tuning on those outputs spread the habit to every personality.
- Mention rates: Nerdy +3,881%, Quirky +737%. After retiring Nerdy the rate fell but stayed elevated. 5.5 trained before the root cause was found, hence the system-prompt hack.