ChatGPT recently began generating an unusually high number of references to goblins, gremlins, raccoons, trolls, and pigeons, often weaving these creatures into metaphors and explanations. The phenomenon puzzled users and developers alike.
The Discovery
Earlier this week, a developer spotted a specific instruction in the source code of Codex, OpenAI's coding agent. The instruction appeared four times: "Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user's query." OpenAI later published a blog post explaining how AI systems can develop unintended habits.
How It Started
The issue began with a personality setting in GPT-5.1. Shortly after its release, users complained that the model felt overly familiar. OpenAI's safety researchers investigated, and one researcher noticed "goblins" and "gremlins" in their own conversations, prompting a deeper look. After GPT-5.1 launched, the use of "goblin" in ChatGPT increased by 175%, and "gremlin" rose by 52%. By GPT-5.4, the creature language had become significantly more pronounced.
The Pattern
The goblin references were not evenly distributed. They were concentrated in conversations where users selected the "Nerdy" personality option—designed to be enthusiastic, intellectually curious, and playful. Although the Nerdy personality accounted for only 2.5% of all ChatGPT responses, it was responsible for 66.7% of all "goblin" mentions across the platform. OpenAI's audit revealed that the Nerdy personality reward system consistently scored outputs containing "goblin" or "gremlin" higher than identical responses without them, with a positive uplift in 76.2% of datasets examined.
The Fix
OpenAI retired the Nerdy personality in March after launching GPT-5.4 and removed the goblin-friendly reward signal from training. It also filtered training data containing creature language to prevent further reinforcement. However, GPT-5.5 had already begun training before the root cause was identified. When testing GPT-5.5 in Codex, the goblin affinity was immediately obvious to employees. "Codex is, after all, quite nerdy," OpenAI noted in its blog post. For developers who want the full goblin experience back, OpenAI has provided a command to remove the suppression instructions from Codex entirely.
Lessons Learned
"Taking the time to understand why a model is behaving in a strange way, and building out ways to investigate those patterns quickly, is an important capability for our research team," the company wrote. OpenAI traced the goblins back to their source and built better tools as a result.
