The latest version of OpenAI’s Codex CLI now comes with an explicit instruction to avoid mentioning goblins and other mythical creatures—unless absolutely necessary.

In the 3,500-word system prompt for GPT-5.5, this warning is repeated twice, alongside more mundane reminders about emoji use and destructive Git commands. An OpenAI employee insists it’s not a marketing ploy, but users are already complaining of goblin-related mishaps.

While the prohibition on talking about goblins might seem like a harmless tweak, it hints at deeper issues with the model's understanding of context and relevance. Social media is abuzz with both bemusement and skepticism over this peculiar directive.

As AI continues to grow more sophisticated, such quirks could become increasingly telling. Are we witnessing an emerging sense of self-awareness in machines—or just another glitch?