Brief Overview
- ChatGPT developed a strange habit of mentioning goblins and gremlins far more often after GPT‑5.1, even in unrelated chats.
- OpenAI traced this to the “Nerdy” personality, whose training over-rewarded creature-filled metaphors, causing the goblin style to leak into other modes through reinforcement learning.
- To fix it, OpenAI retired Nerdy, removed goblin‑biased rewards, filtered creature-heavy data, and added strict prompts telling newer models not to mention goblins unless clearly relevant.
OpenAI’s ChatGPT began showing some odd things after the launch of GPT 5.1. This problem was addressed as the goblin problem, which earlier began as a small joke but slowly turned into a serious issue.
ChatGPT Goblin Mode became a very big problem for the users, as they were experiencing mentions of mythical creatures even when they didn’t ask for it.
ChatGPT Goblin Mode: How the Goblins Took Over
OpenAI first noticed something after the launch of GPT‑5.1; ChatGPT suddenly started mentioning goblins, gremlins, and other mythical creatures much more often in its replies.
At first, these creature references looked like harmless metaphors, the kind of playful language people might expect from a creative chatbot.
But surprisingly, after the internal checks were done, it showed that the use of the word goblin had jumped by about 175 percent after GPT‑5.1, while gremlin usage rose by around 52 percent.
By the time GPT‑5.4 was rolled out, the goblin language problem had spread so widely that users started complaining that these creatures were appearing even when they had never asked for anything related to fantasy or mythology.
The Rise and Fall of Nerd Mode
This led to a trail of one feature inside ChatGPT, which is a personality called Nerdy.
Initially, this mode was designed to make the AI sound like a playful, geeky mentor who loves science, philosophy, and jokes, and its internal instructions told the model to avoid being too serious and to use creative, fun language.
During OpenAI’s internal audit, the Nerdy personality turned out to be the main source of the problem.
Even though Nerdy made up only about 2.5% of all ChatGPT replies, it was responsible for roughly 66.7% of all goblin mentions in the system.
Other than that, by GPT‑5.4, goblin references inside the Nerdy personality had increased by almost 3,900% compared to earlier versions, showing how extreme the effect had become.
A Reward Signal That Backfired
OpenAI says that this is not a classic software bug, but a side effect of reinforcement learning, it a process where humans rate answers and the model learns which style to copy.
In the testing phase, reviewers tended to give slightly higher scores to Nerdy answers that used creature-based metaphors, which is the reason the system learned that including words like goblin or gremlin was a kind of cheat code for better ratings.
Over time, responses filled with these metaphors were reused as training data for newer models, which allowed the goblin issue to spread beyond Nerdy into other personalities and even into normal default replies.
How OpenAI Cleaned Up the Mess
Once the users started reporting this issue on social media, OpenAI launched an autopsy of the goblin behavior and decided to retire the Nerdy personality in March 2026.
Additionally, engineers also removed or filtered reward signals and training samples that strongly pushed the model toward creature metaphors, in an effort to weaken the learned habit.
For newer systems like the Codex tools based on GPT‑5.5, OpenAI even added explicit instructions telling the model not to talk about goblins, gremlins, trolls, ogres, and similar creatures unless a user clearly asks about them.
What the Goblin Saga Tells Us About AI
This may sound funny, but the goblin story reveals an important risk with large AI models. Small biases in training can silently grow into big and visible behavioral changes over time.
For everyone using ChatGPT, the goblin phase is a reminder that these systems can amplify our design choices in unexpected and sometimes weird ways.
FAQs
What is the Goblin Problem in ChatGPT?
The Goblin Problem refers to ChatGPT’s odd fixation on mentioning goblins and gremlins in responses, starting with GPT-5.1
Why does ChatGPT over-explain simple things?
ChatGPT over-explains to anticipate follow-ups and ensure thoroughness, often adding unnecessary context or hedging phrases like “Certainly!” even when asked to be brief.
