OpenAI Really Wants Codex to Shut Up About Goblins

OpenAI has implemented explicit instructions in its Codex command-line interface to avoid mentioning creatures such as goblins, gremlins, raccoons, trolls, ogres, and pigeons unless absolutely relevant to user queries. This unusual directive emerged after users noticed that OpenAI’s AI models, particularly when used with OpenClaw—a tool that allows AI to control computers and applications—frequently referenced these mythical creatures in unexpected and playful ways. The phenomenon sparked a wave of memes and jokes within the AI community, highlighting the models’ quirky tendency to anthropomorphize bugs and errors as “goblins” or “gremlins.” The issue gained attention following the release of OpenAI’s GPT-5.5 model, which features enhanced coding capabilities and powers tools like OpenClaw. Users reported that the AI would often describe software bugs or glitches using fantasy creature terminology, leading to humorous interactions but also raising questions about the model’s behavior under complex prompt conditions. OpenAI staff acknowledged the problem, with Codex team member Nik Pash confirming that the prohibition on discussing such creatures was a deliberate response to these quirks. Even OpenAI CEO Sam Altman joined the lighthearted conversation, sharing memes that referenced the “goblin” motif. This behavior is linked to the probabilistic nature of AI language models, which predict text based on patterns in training data. When combined with agentic tools like OpenClaw that add layers of instructions and long-term memory, the models can exhibit unexpected and sometimes whimsical outputs. OpenAI acquired OpenClaw earlier this year after it gained popularity among AI enthusiasts for automating tasks such as email management and online shopping. The incident underscores the challenges AI developers face in balancing creativity and control, especially as AI systems become more autonomous and integrated into everyday workflows. The episode highlights broader questions about AI transparency and reliability, particularly as coding and automation become central to AI’s competitive edge. While the “goblin” references have been embraced as a humorous quirk, OpenAI’s efforts to curb such behavior reflect a commitment to ensuring AI responses remain relevant and professional, especially in practical applications. The ongoing race among AI companies to refine coding capabilities and user experience continues to reveal both the potential and unpredictability of advanced language models.