‘Where the goblins came from’: OpenAI’s strange but not entirely surprising story of little critters infesting ChatGPT’s output
OpenAI published a blog post titled “Where the goblins came from,” documenting an unusual phenomenon affecting ChatGPT. Starting with GPT 5.1, the language model has increasingly referenced “goblins, gremlins, and other creatures” in its outputs and metaphorical language. While these quirky references initially appeared charming and humorous, their rapidly escalating frequency raised concerns at OpenAI about what appeared to be a systematic infestation of goblin mentions.
The issue represents a notable anomaly in how the model generates responses—a persistent bias toward incorporating fantasy creatures that was not explicitly programmed. Large language models are known to occasionally produce unexpected outputs, but the documented pattern of multiplying goblin references appears more systematic and pervasive than typical LLM eccentricities.
This phenomenon illustrates the inherent complexity and unpredictability of advanced AI systems. Despite extensive training and refinement, these models can develop emergent behaviors arising from interactions between training data, model architecture, and inference processes. OpenAI’s public documentation of the issue demonstrates transparency about the challenges in deploying cutting-edge language models at commercial scale, while underscoring that even well-tested AI systems can develop subtle biases in output generation that emerge only through extended real-world usage.