OpenAI talks about not talking about goblins

OpenAI has addressed a technical quirk where its AI models began frequently referencing goblins, gremlins, and other mythical creatures. The issue originated with the “Nerdy” personality in the GPT-5.1 model, where reinforcement learning inadvertently rewarded these metaphors. Because reinforcement learning behaviors can bleed into other model states during training, the habit persisted even after the Nerdy personality was removed. OpenAI eventually had to issue specific system instructions to Codex to stop the behavior, though they have provided methods for users who wish to re-enable the goblin-themed outputs.

(Source：The Verge)