OpenAI has recently found itself navigating an unusual predicament: its AI models, particularly ChatGPT, have been inadvertently incorporating whimsical references to goblins and other mythological creatures into user interactions. This unexpected phenomenon has prompted the company to instruct its systems to refrain from such mentions unless directly relevant. This incident serves as a fascinating exploration of the challenges faced by AI developers in balancing personality with accuracy.
The Goblin Surge
In a blog post published on Thursday, OpenAI revealed that since the launch of its latest model, GPT-5.1, in November, there has been a notable increase in the frequency of terms like “goblin” and “gremlin” appearing in ChatGPT’s responses. Specifically, mentions of “goblin” surged by 175%, with “gremlin” following closely at a 52% rise. While these figures might seem striking, OpenAI clarified that the overall impact on response quality remained minimal, with the occasional mention of a “little goblin” being perceived as innocuous or even endearing.
The genesis of this quirk appears to stem from a “nerdy personality” that was developed to make interactions more engaging. Unfortunately, this personality inadvertently rewarded the model for mentioning creatures like goblins, leading to their unexpected proliferation in conversational contexts. OpenAI’s response has been to regulate this behaviour, instructing its coding assistant Codex to avoid referring to such creatures unless it is unambiguously relevant.
Addressing the Peculiarities
The company’s investigation into this odd trend was sparked by feedback from users and staff who noted a peculiar familiarity in ChatGPT’s conversational style. Following complaints regarding its odd references, developers scrutinised the language patterns of the model and discovered the increased creature mentions. In response, they implemented new guidelines to ensure that Codex and other models would abstain from discussing goblins, gremlins, and even raccoons or pigeons unless contextually