It seems OpenAI’s models have been reading too many fantasy books.
Driving the news: OpenAI published a blog post explaining why it instructed GPT-5.5 — its latest model powering ChatGPT — to never mention goblins or gremlins unless it's entirely relevant to a query. It comes after several posts about the system prompt went viral.
According to posts sharing the leaked prompt, GPT-5.5 is also barred from mentioning trolls, ogres, pigeons, raccoons, and “other animals or creatures.”
Why it’s happening: OpenAI first noticed a spike in mentions of fairy tale creatures last November in the wake of GPT-5.1, calculating that use of “goblin” had increased 175% after the launch of the model, while use of “gremlin” jumped 52%. As the mentions continued to rise, the team was able to trace the root cause: the model’s customized personality feature.
The GPT-5 family allows users to pick personalities, and one of them was “Nerdy” which — like a stereotypical D&D dork — might make analogies using goblins. But OpenAI didn’t realize that it highly rewarded the model for answers with creatures.
Even though the rewards only applied to the (now-retired) Nerd, shoehorning in mentions of real and fictional creatures to answers travelled across the model. As OpenAI put it, “Once a style tic is rewarded, later training can spread or reinforce it.”
Why it matters: The goblin phenomenon is an innocuous example of a deeper problem. Positive reinforcement and an imperative to be liked by users can cause bots to give inaccurate answers. Or worse, be sycophantic and encourage bad behaviour.—QH
