OpenAI talks about not talking about goblins

OpenAI is opening up about its goblin problem. After report from Wired revealed instructions to the OpenAI version of the document to “not talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures,” the introduction of AI produced an explanation on his website, calling the creation “strange practice” his examples created as a result of their studies.

As explained in the blog post, OpenAI started to recognize metaphors about wolves and other creatures. starting with its GPT-5.1 version – especially when using the “Nerdy” personality. OpenAI says that the problem continued to grow with the release of successive models, until it found that its reinforcement learning provided models similar to Nerdy’s personality, from which new models were trained.

The rewards were used in the Nerdy culture, but reinforcement learning does not guarantee that the learned behaviors are as good as they are made out to be. Once the structure is rewarded, subsequent training can be generalized or reinforced, especially if the results are used in supervised or preferred training.

Although the talk of goblins and gremlins stopped after OpenAI abandoned Nerdy’s persona in March, it hasn’t completely disappeared. and GPT-5.5 within its Codex scripting tool, where OpenAI began training the model before finding a “trigger.” The company had to give Codex specific instructions not to talk about the mythical creatures. But if you’d rather have your own AI code with some goblin sprinkled in, OpenAI has it covered he shared a way to change his instructions.