OpenAI’s CEO Sam Altman finds himself wrapped in courtroom intrigue, while OpenAI engineers have addressed a minor glitch with ChatGPT. A comprehensive report regarding goblins was published on OpenAI’s website. Following the introduction of GPT 5.1, ChatGPT models exhibited an unusual obsession with goblins and gremlins, frequently referencing them in unrelated situations. Users brought this to OpenAI’s attention, prompting an investigation. The root of the issue was traced back to a misinterpretation of “nerdy” behavior.

ChatGPT provides different “personalities” to customize the tone of responses. One such personality, “nerdy,” was recently discontinued. Even though “nerdy” responses constituted merely 2.5% of ChatGPT’s interactions, an overwhelming 66.7% of goblin references originated from this category. These responses began to infiltrate other personality types outside of “nerdy.”

OpenAI clarified that reinforcement learning can lead to the propagation of learned behaviors beyond their initial context. Consequently, OpenAI phased out the “nerdy” personality in March and directed models to refrain from discussing goblins or gremlins unless fitting.