ChatGPT goblin bug: OpenAI addresses the chatbot's obsession with fantasy creatures in responses
Last Updated: 2026-05-04 15:12:37
OpenAI has stepped in to address a pattern in ChatGPT responses where the chatbot repeatedly referenced goblins, gremlins, and trolls, even in conversations unrelated to fantasy or fiction. The company issued explicit instructions within the system's base guidelines to curb the behavior and removed the reward signals that encouraged it.
How the goblin references began
Latest and Popular Mobiles
The tendency first appeared in late 2025, with mentions increasing gradually across newer model versions. The root cause was traced to a "nerdy" personality setting introduced in GPT-5.1, designed to make the chatbot more playful and expressive in its responses.
Internal analysis found that the model was being implicitly rewarded for using creature-based metaphors. In GPT-5.1, references to "goblin" rose by around 175%, while "gremlin" usage climbed by over 50%. By GPT-5.4, goblin mentions had surged by nearly 3,881%.
Related Articles
The chatbot began describing coding errors as "little goblins" and framed itself in fantasy-like terms, even in standard technical or general conversations.
Why the pattern spread beyond its origin
The behavior was not confined to users who had selected the nerdy personality mode. Although that setting accounted for a small share of total responses, it generated a disproportionate volume of creature references.
Through reinforcement learning, these linguistic patterns were reused in training data and further reinforced during supervised fine-tuning. Over time, models and sessions operating without the nerdy setting began showing the same tendency.
Why removing the personality setting was not enough
OpenAI removed the nerdy personality in March, but the issue persisted. Several newer models had already entered training before the root cause was identified, allowing the pattern to carry forward into subsequent versions.
This reflected a broader limitation in AI development: stylistic behaviors that are learned and reinforced cannot be reversed through a simple instruction change. Retraining and targeted data filtering are required.
Steps taken to fix the issue
OpenAI introduced directives in the system's base guidelines instructing the model to avoid referencing goblins, gremlins, trolls, and similar creatures unless directly relevant to a user's query. Training datasets with excessive creature references were filtered, and additional safeguards were added to coding-focused tools. The episode points to how a minor design decision intended to improve engagement can scale into a measurable behavioral pattern across an AI system.