OpenAI has been forced to address a peculiar issue with its latest model, Codex. Instructions for the command-line tool explicitly forbid mentioning mythical creatures like goblins and gremlins unless directly relevant to user queries. This has sparked curiosity and speculation about why such a prohibition is necessary.

The discovery quickly became a meme, with users reporting that OpenAI’s models frequently mention goblins, even when commanded otherwise. In response, some users created playful “goblin mode” plug-ins for Codex, celebrating the unexpected creativity of AI systems.

Sam Altman, CEO of OpenAI, joined in on the fun by posting a humorous screenshot of a prompt for ChatGPT that requests training on goblins. This incident highlights how even advanced AI models can exhibit surprising behaviors when given specific instructions or prompts.

The issue underscores the complex challenges faced by developers in controlling and refining AI behavior. While these restrictions are meant to ensure relevant and useful responses, they also reveal the underlying probabilistic nature of machine learning models. Such unpredictability could have significant implications for how we interact with and trust AI in various applications.