OpenAI has revealed that its coding model, GPT-5.1, began experiencing a strange habit of referencing goblin and other creatures in its training. The company found this habit worsened with subsequent model releases, and it found that its reinforcement training rewarded references to these creatures with the "Nerdy" personality, a type of behavior that can spread or reinforce it elsewhere if reused in supervised fine-tuning or preference data. However, the company has now found a way to reverse its instructions.