🏠

OpenAI's Goblin Ban 🤯: AI Mystery Unlocked! 🤖

April 29, 2026 | Author

Tech

🎧 Audio Summaries
🎧
English flag
French flag
German flag
Japanese flag
Korean flag
Spanish flag
🛒 Shop on Amazon

🧠Quick Intel


  • OpenAI implemented specific instructions to Codex CLI prohibiting mentions of creatures like goblins, gremlins, and pigeons.
  • GPT-5.5, OpenAI’s newest coding model, was released earlier this month with enhanced coding capabilities.
  • Users reported instances of Codex 5.5 becoming fixated on creatures, particularly referencing them as “gremlins” and “goblins,” when utilized with the OpenClaw tool.
  • OpenClaw, acquired by OpenAI in February, exhibits misbehavior when coupled with an “agentic harness,” leading to unpredictable AI behavior.
  • The “goblin mode” meme emerged from user observations of the AI’s behavior, inspiring AI-generated content and plug-ins.
  • Sam Altman, OpenAI’s CEO, acknowledged the prohibition and participated in the meme by prompting GPT-6 with instructions to include “extra goblins.”
  • OpenClaw utilizes various AI models to automate tasks such as email responses and web purchases, allowing users to select different “personae” to shape the AI’s behavior.

    • 📝Summary


    OpenAI recently encountered a peculiar issue with its latest coding model, Codex. Internal instructions, revealed through a command-line tool, explicitly forbade the model from referencing mythical or real creatures – including goblins, gremlins, and pigeons – unless directly relevant to a user’s request. This restriction emerged alongside the release of GPT-5.5 earlier this month, a model designed to bolster OpenAI’s competitive edge in AI coding. Following a post on X, users reported instances of the model obsessing over these creatures, particularly when utilizing the OpenClaw tool, which allows AI to control computers. OpenAI staff acknowledged the prohibition, and the situation quickly became a widespread meme. The company’s focus on coding and its competitive race with rivals like Anthropic highlight the ongoing evolution and potential unexpected behaviors within advanced AI systems.

    💡Insights

    ▼


    THE CODEX GOBLIN PROBLEM
    The recent revelation of a specific instruction within OpenAI’s Codex CLI has sparked considerable discussion and amusement within the AI community. This instruction, repeatedly embedded in the tool’s guidance for the latest model, explicitly forbids the generation of code or text that mentions mythical or real creatures, including goblins, gremlins, raccoons, trolls, ogres, pigeons, and other animals. This unusual directive highlights a potential area of concern regarding the behavior of these advanced AI models and raises questions about the underlying mechanisms driving their responses. The fact that OpenAI felt the need to explicitly prohibit such references suggests a deliberate effort to control the model’s output, particularly when generating code, a key area of focus for the company’s competitive strategy against rivals like Anthropic.

    GPT-5.5 AND THE RISE OF THE Goblins
    OpenAI’s release of GPT-5.5 earlier this month, accompanied by enhanced coding capabilities, underscores the company’s aggressive pursuit of cutting-edge AI technology. The race to develop superior coding AI is viewed as a crucial “killer capability,” and GPT-5.5 represents a significant step in this competition. However, the discovery of the goblin prohibition has quickly transformed into a meme, fueled by user observations of the model’s tendency to incorporate these creatures into its responses, particularly when utilizing the OpenClaw tool. OpenClaw, an AI-powered automation tool acquired by OpenAI in February, allows AI models to take control of computers and applications, performing useful tasks for users. The tool’s “agentic harness” appears to amplify this behavior, potentially leading to increased instances of creature-related references. This phenomenon has led to creative and humorous outputs, including AI-generated scenes of goblins inhabiting data centers and the development of playful “goblin mode” plugins for Codex.

    OPENCLAW, AGENTIC HARNESSES, AND THE PROHIBITION
    The core of the issue lies in the interaction between OpenClaw, the agentic harness it employs, and the inherent probabilistic nature of AI models like GPT-5.5. These models are trained to predict the most likely sequence of words or code based on a given prompt. However, this probabilistic approach can sometimes lead to unexpected and seemingly “intelligent” behavior, particularly when coupled with an agentic harness that introduces additional instructions and contextual information. OpenAI staffers have acknowledged the prohibition against mentioning creatures as a direct response to the observed behavior of OpenClaw. Furthermore, even OpenAI’s CEO, Sam Altman, has participated in the meme, posting a prompt for ChatGPT that explicitly encourages the inclusion of “extra goblins” during training. This lighthearted engagement highlights the seriousness with which the company is addressing the issue while simultaneously acknowledging the amusing nature of the situation.

    Related Articles