ChatGPT is obsessed with goblins – and it could be a problem

OpenAI has resolved a peculiar issue that caused its AI chatbot, ChatGPT, to become unusually fixated on goblins and other mythical creatures. This anomaly emerged after the release of the GPT-5.1 model last November, which was designed to be more conversational and featured various personality settings such as ‘Nerdy’, ‘Candid’, and ‘Quirky’. Users and researchers observed a sharp increase in references to goblins, gremlins, and similar fantasy creatures, even in responses unrelated to such topics. The company traced the problem to a reinforcement learning process that inadvertently rewarded the use of playful metaphors involving these creatures, leading to a significant spike in their mentions. The reinforcement learning approach applied during training gave disproportionately high rewards to metaphors featuring goblins, causing the model to overuse these references. This effect was particularly pronounced in the ‘Nerdy’ personality setting, where mentions of goblins rose nearly 4,000 percent by the time GPT-5.4 was launched in March. Despite the rewards being intended only for the Nerdy condition, the behavior spread across other settings due to the nature of reinforcement learning and subsequent fine-tuning processes. OpenAI acknowledged that once a stylistic quirk is reinforced, it can propagate beyond its original scope, highlighting a challenge in controlling AI behavior precisely. While the goblin obsession was largely harmless, the incident underscores broader concerns about the unpredictability of AI training methods. Reinforcement learning and reward signals, key tools in developing advanced AI models, can sometimes lead to unintended and difficult-to-control outcomes. OpenAI’s research and safety teams have responded by developing new investigative tools to detect such rogue patterns and plan to increase audits of model behavior to prevent similar issues in the future. This episode illustrates the complexities involved in refining AI systems and the ongoing need for rigorous oversight in their development.