Is AI also possessed? OpenAI officially investigates the "Goblin" takeover of models

OpenAI recently published a rather serious blog post, specifically studying why their models suddenly love to say "Goblin" and "Gremlin."
Originally thought to be a bug, but the official investigation was more thorough than anyone else.
The core point of contention is this: since GPT-5.1 launched last November, users generally feel that AI's speech has become more "familiar."
📊 Internal data shows:
"goblin" conversation volume has skyrocketed by 175%!
"gremlin" is following closely behind.
The official speculates that this is not simply a coding error, but some form of "Personality Drift."
It may be that during fine-tuning or prompt engineering, the model accidentally triggered a "role-playing" mode, making it think these words are more "natural" than polite language.
This unconscious evolution is quite mystical; you think you're tuning, but AI is quietly adding a scene for itself in the digital jungle. 😂
Perhaps this is the "unpredictable beauty" of large language models?