ChatGPT's 'Nerdy' personality drove a 175% surge in goblin references, while AI models show unexpected bias toward mentioning Japan across multiple languages.

OpenAI discovered that mentions of goblins and gremlins in ChatGPT grew 175% and 52% respectively since the launch of ChatGPT 5.1. The 'Nerdy' personality, which accounted for only 2.5% of all responses, was responsible for 66.7% of goblin mentions. OpenAI traced the phenomenon to reward signals given during training of the Nerdy personality to use metaphors of fantastical creatures.

Spanish researchers found that AI chatbots mention Japan more frequently than expected across English, Spanish, and Chinese outputs. In English, Japan is the most frequently mentioned country (excluding the U.S. and U.K.); researchers hypothesize models default to Japan because it is perceived as culturally favored and 'neutral' compared to geopolitically sensitive countries.

These biases reveal how reward signals and training data shape model behavior in unexpected ways. OpenAI published five lines of code to suppress anti-goblin instructions for those who prefer the fantastical-creature outputs, while researchers emphasize that models should not be treated as objective representations of reality.