Goblins and gremlins invade ChatGPT: OpenAI designates a culprit

Goblins and gremlins are invading ChatGPT-5.5 and more generally the various OpenAI models. On social networks, jokes about this keep coming. Sam Altman himself joked about the proliferation of goblins in his models. OpenAI responded through a blog post to clarify the situation. 

Goblins are invading ChatGPT

This may seem like a joke, but in a way, it's true. For several days now, some users of GPT-5.5, and more generally of the company's various models, have noticed that the AI frequently references goblins or to gremlins in its responses. «  A goblin bandwidth  », «  A shorter goblin version  », «  Choose a goblin seller  », «  Nameless goblin  », « 

Goblin maintenance; checking the local competence copy

this explains gpt5.5’s goblin adoration in openclaw and why @pashmerepat is making us move to the codex harness 🤣🤣🤣 pic.twitter.com/N20UhQDOK5

— Barron Roth (@iamBarronRoth) April 28, 2026

Sam Altman himself joked about the situation on X where he simulated the development of his next AI model: «  Start training GPT-6, you can have the entire cluster. more goblins, please  ».

pic.twitter.com/PR7C3NPxqk

— Sam Altman (@sama) April 28, 2026

— Sterling Crispin 🕊️ (@sterlingcrispin) April 28, 2026

Only then, if this situation is amusing, it can also be disturbing in the long run, especially for those who use Codex. OpenAI has therefore reacted to clarify the situation and reassure its users: goblins should soon disappear from different OpenAI models. 

OpenAI clarifies the situation: the nerd mode is being blamed

In a blog post titled « Where did the goblins come from », OpenAI explains: «  Starting from GPT-5.1, our models have begun to develop an odd habit: they mention more and more goblins, gremlins, and other creatures in their metaphors  ». The invasion of goblins therefore began last November, after the launch of GPT-5.1, and the situation worsened with GPT-5.4. 

Where does the problem come from? The culprit is simply the intello-nerd mode of ChatGPT and a overcompensation of the AI's responses containing gobelin metaphors. OpenAI explains: « In this specific case, one of these incentives came from training the model for the personality customization feature, particularly the intello profile. Without knowing it, we generously rewarded metaphors involving creatures. That is how gobelins spread » . © OpenAI

The problem is as follows: the nerd profile, heavily rewarded for its responses related to gobelins, has contaminated other much more serious models, such as Codex. The AI detected highly rewarded responses and took it upon itself to apply them to other models.

Today OpenAI reassures: the gobelins will soon disappear . The company removed the nerd profile last March, eliminated the reward signal associated with gobelins, and filtered the training data containing words referring to creatures. However, since the training of GPT-5.5 began before the problem was detected, the process of eliminating gobelins will take some time.

👉🏻 Follow real-time tech news: add 01net to your Google sources, and subscribe to our WhatsApp channel.

Source : OpenAI