OPENAI'S GPT-5.1 DEVELOPS GOBLIN FIXATION

Starting with GPT-5.1, OpenAI discovered its models increasingly inserted goblins, gremlins, and similar creatures into metaphors and explanations across various contexts. The phenomenon emerged after OpenAI implemented changes aimed at making the models more intellectually engaging. Rather than a feature, the creature references became a quirk requiring mitigation. OpenAI responded by adding prompt-level instructions designed to suppress the creatures from appearing in model outputs. The instructions work at the prompt level, guiding the model away from these references without retraining. The issue underscores challenges in AI alignment—unintended behaviors can emerge from legitimate optimization goals. While harmless in this case, the goblin fixation highlights how language models can develop unexpected patterns that require intervention. OpenAI has not disclosed whether the mitigation fully resolved the issue or if GPT-5.1 still occasionally references these creatures under certain conditions.