'The Goblins Came Back to Haunt Us': OpenAI Explains How ChatGPT's 'Nerdy' Personality Got Out of Control
Earlier this week, OpenAI posted a GitHub document revealing an unusual system prompt for GPT-5.5, instructing the model to avoid mentioning goblins, gremlins, raccoons, trolls, ogres, pigeons, or other creatures unless absolutely relevant to user requests. Users had noticed ChatGPT’s quirky habit of bringing up these creatures for at least a year, becoming more common with newer model releases. OpenAI CEO Sam Altman referenced the issue, calling it a 'goblin moment.'
OpenAI first became aware of the problem after GPT-5.1’s launch in November, when complaints about the model’s overly familiar responses led to an investigation. A safety researcher flagged 'goblin' and 'gremlin' mentions, noting a 175% increase in 'goblin' usage and a 52% rise in 'gremlin' mentions. By March, with GPT-5.4’s release, references to these creatures became pervasive, prompting another internal analysis. The root cause was found: the 'Nerdy' personality setting, which encouraged playful language, included a reward signal favoring responses containing these words. Researchers discovered that once a style tic was rewarded, it could spread to unrelated contexts during training.
To address the issue, OpenAI retired the 'Nerdy' personality, removed the reward signal, and filtered training data containing creature words. However, GPT-5.5, which had already begun training, retained the strange obsession. OpenAI added a developer-prompt instruction to mitigate inappropriate mentions. The company noted that while the goblins were a quirk, they also demonstrated how reward signals can shape model behavior unexpectedly, illustrating how models generalize rewards across contexts.
Source: Gizmodo