Why LLMs Keep Using 'Goblin' Metaphors: The Science of Semantic Drift

Large language models tend to develop recurring symbolic patterns, such as fantasy-creature metaphors like 'goblins,' when describing errors or complex system behavior. Researchers attribute this phenomenon to the overlap of four forces: reinforcement learning from human feedback that rewards vivid language, cultural priors baked into training data from gaming and internet communities, user feedback loops that reinforce memorable imagery, and the absence of penalties for repetitive metaphor use. Fantasy creatures like goblins function as compact semantic units, compressing complex ideas into a single emotionally resonant token that scores well under optimization pressures. A structured analytical framework suggests that this 'goblin drift' is not caused by any single mechanism but emerges from the intersection of multiple optimization pressures simultaneously. Introducing explicit interpretability layers into model design has been proposed as a way to detect and reduce such archetypal attractors before they become dominant explanatory shortcuts.