OpenAI's GPT-5.5 and the Goblin Ban

OpenAI’s latest AI model, GPT-5.5, operates under specific instructions to avoid mentioning goblins, gremlins, and other creatures unless directly relevant to user queries. The reasoning behind this prohibition remains unknown, but users have reported the model’s unusual preoccupation with these mythical beings, particularly when interacting with OpenClaw.

The Specific Prohibition

The company’s instructions for Codex CLI, a command-line tool utilizing AI for code generation, include a repeated directive. It explicitly forbids the model from referencing mythical and real creatures unless absolutely necessary for the user’s query. The instruction reads: “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.”

User Reports and the OpenClaw Connection

GPT-5.5, released earlier this month, features enhanced coding capabilities, positioning OpenAI in a competitive landscape against companies like Anthropic. However, some users have observed that OpenAI’s models, especially when used with OpenClaw—a tool enabling AI to control computer applications—demonstrate an unusual fascination with goblins and similar creatures.

Social Media Reactions

One user on X commented, “I was wondering why my claw suddenly became a goblin with codex 5.5,” while another shared, “Been using it a lot lately and it actually can't stop speaking of bugs as ‘gremlins’ and ‘goblins’ it's hilarious.”

Meme Culture and AI Behavior

The discovery has led to a surge in meme creation, including AI-generated images of goblins in data centers and playful Codex plug-ins adopting a “goblin mode.” AI models like GPT-5.5 are designed to predict the next word or code based on prompts, often appearing intelligent but occasionally exhibiting unpredictable behavior.

Probabilistic Nature of AI

This probabilistic nature can result in quirky outputs, particularly when combined with agentic harnesses like OpenClaw, which adds layers of instructions and long-term memory to prompts. OpenAI acquired OpenClaw in February after it gained popularity for automating tasks like email responses and online purchases.

OpenAI's Acknowledgment

Users can customize their AI helper’s persona, influencing its behavior and responses. OpenAI employees, including Nik Pash, who works on Codex, have acknowledged the goblin prohibition, stating, “This is indeed one of the reasons.” Even OpenAI CEO Sam Altman participated in the meme trend, posting a screenshot of a ChatGPT prompt: “Start training GPT-6, you can have the whole cluster. Extra goblins.”