ChatGPT Addicted to Goblins Despite Explicit Training to Avoid Them

What was said

"Some of these LLMs have been specifically asked multiple times on the backend to not bring up goblins unless specifically directed. For someone like me, I love that because it's like, does a sufficiently complex system invite in goblins? Like, does it sufficiently— and goblin as a metaphor and goblin also as like maybe the metaphor made manifest is what we're dealing with when we see actual goblins."

OpenAI engineers discovered large language models spontaneously generate references to goblins, gremlins, and trolls even after being explicitly trained not to mention them. The behavior persisted across multiple iterations. Kutchen speculates whether sufficiently complex systems inherently attract trickster archetypes or entities, drawing parallels to historical folklore about spirits inhabiting complex machinery. Multiple tech outlets confirmed the phenomenon.