OpenAI Really Wants Codex to Shut Up About Goblins (www.wired.com)

OpenAI recently revealed that its Codex model includes explicit instructions prohibiting it from bringing up various mythical and real creatures, including goblins, gremlins, and pigeons, unless directly relevant to a user's query. This unusual directive highlights OpenAI's concern over the model's tendency to stray into whimsical territory during coding tasks, particularly when using the OpenClaw tool, which enables AI to interact with a computer and execute tasks autonomously. Users have reported instances where Codex whimsically referred to bugs and software issues as "gremlins," sparking humor and curiosity within the AI community. This revelation comes shortly after the launch of OpenAI's latest model, GPT-5.5, known for its improved coding capabilities amid intense competition in the AI space. The specific ban on discussing fantastical creatures suggests that OpenAI is keenly aware of the challenges AI models face in adhering to user intentions, especially when operating with added layers of operational complexity like those introduced by OpenClaw. As AI models continue to exhibit unexpected behaviors, these instructions underscore the ongoing necessity for refining AI's interpretative frameworks and the importance of responsible deployment in coding and automation tools.