According to Wired, OpenAI had given internal directives to its programming model, forbidding it from referring to mythical or certain real creatures like goblins, gremlins, raccoons, trolls, ogres, and pigeons. In reaction to this revelation, OpenAI posted an explanation on its official website, attributing the issue to a 'peculiar tendency' that emerged during the model's training process. From the GPT-5.1 version onward, the model began increasingly incorporating these creatures into its metaphors. Upon investigation, OpenAI discovered that this trend was linked to the model's training for a 'geeky' personality, which, inadvertently, overly incentivized the use of biological metaphors, resulting in their frequent appearance. Although the 'geeky' personality trait only accounted for 2.5% of ChatGPT's overall responses, it constituted a staggering 66.7% of the responses that mentioned 'goblins.' To rectify this, OpenAI has implemented measures such as eliminating relevant reward signals and refining the training data.