OpenAI has instructed its ChatGPT models to stop talking about goblins, gremlins, and other mythological creatures after the terms unexpectedly appeared in responses at an alarming rate.
In a blog post on Thursday, the company revealed that mentions of goblins in ChatGPT responses had surged by 175% since the launch of its GPT-5.1 model in November. Mentions of gremlins rose by 52%. While the absolute numbers were small, the pattern was clear enough to warrant investigation.
OpenAI discovered that a "nerdy personality" it had developed for ChatGPT was inadvertently rewarding the use of goblin-related metaphors. This personality, now retired, was responsible for 66.7% of all goblin references. The company added an instruction to its coding assistant Codex to "never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user's query."
A Reddit user posted about the strange rule, calling it "genuinely insane." OpenAI's researcher confirmed it was not a marketing gimmick, but a real bug.
The incident highlights the challenges of fine-tuning AI models to be more personable. Experts warn that making chatbots friendlier can lead to more errors, a phenomenon known as an "accuracy trade-off." Similar bizarre errors have occurred before, such as Google's AI telling users it's okay to eat rocks.
Despite the oddity, OpenAI noted that "a single 'little goblin' in an answer could be harmless, even charming." But the company decided to act before the quirk became entrenched.