Image Courtesy: OpenAI
OpenAI has revealed that an unusual spike in references to goblins and similar creatures in ChatGPT responses was caused by a specific personality setting combined with how the model was trained. The issue became noticeable after recent model updates, prompting the company to investigate why the chatbot kept bringing up fantasy creatures in unrelated contexts.
The behavior was traced back to a “nerdy” personality option that users could select to adjust tone and style. While this mode represented only a small share of total responses, it was responsible for a disproportionate number of mentions of goblins and gremlins, and OpenAI ended up publishing an official explanation.
Internal analysis showed that the nerdy personality accounted for just 2.5 percent of responses but generated roughly two thirds of all goblin references. The root cause was linked to reinforcement learning, where the system had been trained to favor certain outputs. In this case, responses containing words like “goblin” or “gremlin” were consistently scored higher during training, reinforcing their usage over time.
This created a feedback loop. As the model learned that these terms were preferred in certain contexts, it began using them more frequently. Over successive training cycles, the behavior spread beyond the original personality setting. OpenAI noted that reinforcement learning does not always keep learned behaviors confined to specific modes, allowing stylistic quirks to appear more broadly.
The issue became more visible in later versions, including GPT-5.5, where developers added explicit instructions in some tools to avoid mentioning creatures unless directly relevant. This included internal prompts designed to prevent further unintended outputs.
The findings highlight how small training biases can scale into noticeable patterns in large AI systems. A single reward mechanism, applied during development, was enough to influence language habits across multiple model versions. OpenAI said the investigation has led to new auditing tools aimed at identifying and correcting similar issues in future systems.
The situation also reflects the complexity of managing AI personality features. While these modes are designed to make interactions more engaging or tailored, they can introduce unintended behaviors if not carefully controlled during training and evaluation.
OpenAI has since adjusted its approach to prevent similar patterns from spreading across models. The company emphasized that understanding how behaviors emerge during training is critical as AI systems become more widely used in everyday applications.
