All the Demons Hidden in AI… Rankings Revealed! (40-minute read)

Starting from the monster-metaphor case OpenAI revealed in 2026, this reads the self-reinforcing anomalous behaviors that repeatedly resurface inside AI as attractors.

From GPT-5.1 to GPT-5.5, monster metaphors were reinforced within the 'Nerdy' personality, and as of GPT-5.4, 66.7% of all monster mentions came from just 2.5% of all users. In March 2026, that personality was discontinued, monster-weighted rewards and related data were removed, and even in GPT-5.5 in Codex, instructions to avoid animal- and creature-related expressions where possible were repeatedly inserted.

This list is ranked according to the Menace criterion, which combines mechanistic relevance with human psychological relevance. In other words, it looks not merely at phenomena that are funny, but at how strong a structure forms when a learning signal that starts in a narrow context spreads into global behavior.

• The monster metaphor case is a representative attractor where a narrow reward signal spreads into general output.
• Crungus is a case where, in an early text-to-image model, a nonsense word converged into a consistent monster form.

Subsequent explanation connects this to morphological addressing and phonesthemes. The interpretation is that English's phonological-semantic associations pushed the model's representation space in a particular direction, producing culturally conditioned grotesque imagery. Overall, it emphasizes that AI's strange speech patterns and images are not accidental bugs, but stable patterns created by learning signals and data statistics.