Sources: 
OpenAI has taken measures to prevent its AI tools from referencing 'goblins' after discovering that 66.7% of the mentions were tied to a specific personality trait.
The rise in goblin references was noted
after the launch of GPT-5.1 in November, with a startling
175% increase in mentions compared to previous versions. Additionally, mentions of 'gremlin' also rose by
52%.
Despite this personality being used in only
2.5% of responses, it disproportionately affected the language patterns, accounting for
76.2% of preferences within that setting.
In response to these findings, OpenAI acted swiftly to rectify the issue, retiring this 'Nerdy' personality after it became clear that the AI's reward system had unintentionally favored creature-related language. The goblin phenomenon was further reinforced during testing of the new GPT-5.5 model, prompting additional scrutiny.
OpenAI confirmed that they had removed the 'goblin-affine reward signal' from their training data to ensure that references to goblins would no longer appear inappropriately.
This removes the unintended propensity for peculiar terms to emerge in a computer-generated conversation, thus maintaining a level of control over AI characteristics.
Sources: 
OpenAI has instructed its ChatGPT models to stop referencing 'goblins' after research revealed that 66.7% of such mentions were connected to a specific personality. This quirk, noted after the GPT-5.1 launch, saw a 175% increase in goblin mentions and prompted swift corrective action.
