OpenAI is addressing its goblin issue. After a Wired report disclosed instructions to its coding model to avoid discussing goblins, gremlins, raccoons, trolls, ogres, pigeons, or similar creatures, the AI startup provided an explanation on its website, referring to these references as a “strange habit” its models developed from their training.
In the blog post, OpenAI noted that metaphorical references to goblins and other creatures began with the GPT-5.1 model, particularly using the “Nerdy” personality option. OpenAI found the problem worsened with subsequent models, as reinforcement training rewarded the quirky metaphors associated with the Nerdy personality.
Although goblin and gremlin mentions decreased after OpenAI removed the Nerdy personality in March, they persisted in GPT-5.5 inside its Codex tool, as the model was trained before identifying the “root cause.” To counter this, OpenAI gave Codex specific commands to avoid these mythological creatures. However, if you wish to incorporate goblin themes into your AI coding, OpenAI has shared a method to reverse these instructions.