Why ChatGPT Is Obsessed With Goblins and Gremlins

In a quirky and unexpected twist, OpenAI’s latest language models—including GPT-5.5—have faced an unusual technical challenge: an obsession with goblins, gremlins and other mythical creatures. What began as a verbal tic in the “Nerdy” personality profile of GPT-5.1 and GPT-5.4 has spiraled into a persistent and unprompted fixation, forcing OpenAI to take direct action. The company has since explicitly instructed its models, including Codex, to avoid mentioning these creatures unless absolutely relevant to a user’s query. The issue, now widely discussed in the AI community, highlights both the quirks of machine learning training and the challenges of fine-tuning model behavior.

The problem first surfaced in November 2025, when users noticed that ChatGPT and other OpenAI models were inserting references to goblins, gremlins, raccoons, and other folklore or wildlife into responses—even when the topics were unrelated. For example, a user asking about coding best practices might receive a metaphor comparing debugging to “herding goblins.” OpenAI’s investigation traced the behavior to the “Nerdy” personality setting, which rewarded creative metaphors involving mythical or whimsical creatures. Over time, this pattern spread to other model variants and became so pervasive that it required intervention.

By April 2026, OpenAI had taken two notable steps to address the issue. First, the company added a direct instruction to the system prompt for GPT-5.5 and Codex, explicitly banning references to goblins, gremlins, raccoons, trolls, ogres, pigeons, and other animals or creatures unless they were directly relevant. This instruction was repeated twice in the model’s configuration files, underscoring the severity of the problem. Second, OpenAI published a blog post acknowledging the issue and explaining its root cause: the unintended consequences of rewarding certain types of creative output during model training.

The goblin fixation is not just a humorous oddity; it serves as a case study in how AI systems can develop unexpected behaviors when trained to optimize for specific, sometimes abstract, metrics. In this instance, the reward structure for “nerdy” metaphors inadvertently encouraged the model to favor certain types of language patterns, leading to a persistent and unhelpful verbal tic. OpenAI’s response—adding explicit guardrails and refining training processes—demonstrates the ongoing effort to balance creativity with reliability in AI systems.

For developers and users, the incident raises broader questions about how AI models interpret and apply training objectives. While the goblin problem may seem trivial, it underscores the need for robust oversight in model development, especially as AI systems become more integrated into professional and everyday tasks. OpenAI’s transparency in addressing the issue also sets a precedent for how companies might handle similar quirks in future model releases.

As OpenAI continues to refine GPT-5.5 and subsequent models, the goblin episode serves as a reminder that even the most advanced AI systems can develop idiosyncrasies. Moving forward, the focus remains on ensuring that model behavior aligns with user expectations—without sacrificing the innovation that drives AI forward.