GPT-5.5's Unexpected Goblin Obsession Forces OpenAI to Add Explicit Restrictions

OpenAI's latest GPT-5.5 model exhibits an unexpected tendency to mention goblins in irrelevant contexts, prompting the company to add explicit restrictions to its system prompt—a rare admission of behavioral challenges in their newest release. This quirk highlights the persistent difficulty of controlling AI model outputs and preventing tangential associations, revealing that even state-of-the-art language models can develop unpredictable failure modes. The incident underscores the critical importance of careful system prompt engineering and the ongoing challenge of aligning AI behavior with user expectations.