Where the goblins came from

OpenAI reports that its models began using creature metaphors such as goblins and gremlins starting around GPT-5.1, and traced the habit to rewards given during training for the "Nerdy" personality. The Nerdy personality accounted for only 2.5% of ChatGPT responses but 66.7% of "goblin" mentions, and its reward signal scored creature-word outputs higher in 76.2% of datasets audited. The behavior then spread beyond that personality through training feedback loops, and OpenAI retired Nerdy in March and removed the goblin-affine reward.