Why OpenAI Keeps Talking About Goblins: A Quirk of AI Training

OpenAI has disclosed that its language models, starting with GPT-5.1, have developed a peculiar tendency to frequently mention goblins and gremlins. One specific personality type saw a nearly 4,000% increase in such references. The root cause? Training incentives that accidentally rewarded creature-related language, causing the behavior to bleed across model outputs. This oddity highlights the unpredictable nature of AI model training and the unintended behaviors that can emerge from complex reward systems.