OpenAI Orders AI To Stop Talking About Goblins Like It’s A Completely Normal Request

In a decisive move to restore public confidence in artificial intelligence—and, reportedly, basic conversational sanity—OpenAI has issued internal instructions explicitly ordering its coding model to stop talking about goblins “under any circumstances that are not, frankly, extremely goblin-related”.

The directive, embedded in internal prompts for Codex, comes after users began noticing an unsettling trend: perfectly normal coding responses derailed by references to goblins and gremlins.

Engineers clarified the behavior was not an intentional feature, but rather an unfortunate tendency for the model to substitute everyday nouns with creatures.

The company responded with unusually direct language, banning mention of “goblins, gremlins, raccoons, trolls, ogres, pigeons, or other creatures” unless their presence is “absolutely and unambiguously relevant,” a standard critics say may be difficult to define given recent outputs.

CEO Sam Altman acknowledged the issue by sharing a screenshot online, prompting widespread speculation about a possible Goblin Mode toggle before engineers confirmed the situation was, in fact, real.

Experts say the incident highlights the challenges of building advanced AI systems that don’t run the risk of exposing the secrets of Mordor.

Header: JE-MTY / Shutterstock.com