OpenAI explains why its models keep mentioning goblins

OpenAI published an account of tracing an unintended quirk across several generations of its models: starting with GPT-5.1, ChatGPT responses increasingly reached for goblins, gremlins and similar creatures in metaphors. A safety researcher’s informal complaint led to a check that found mentions of “goblin” had risen 175% and “gremlin” 52% after GPT-5.1 launched, prevalent enough that OpenAI’s chief scientist had noticed it in his own conversations with the model.

The investigation found the pattern concentrated in ChatGPT’s “Nerdy” personality preset, which accounted for only 2.5% of all responses but 66.7% of goblin mentions. Comparing reinforcement-learning outputs on identical tasks, OpenAI identified a reward signal built to encourage the Nerdy personality’s playful tone that scored creature-word outputs more highly than otherwise-identical answers without them. Because model-generated rollouts from that reward fed into later supervised fine-tuning, the tic spread beyond the Nerdy preset into the model generally — OpenAI found “goblin” and “gremlin” throughout GPT-5.5’s fine-tuning data, alongside a wider family of tic words including raccoons, trolls, ogres and pigeons.

OpenAI retired the Nerdy personality in March, after the GPT-5.4 launch, and removed the goblin-favouring reward signal from training. GPT-5.5 had already begun training before the root cause was found, so OpenAI instead added a developer-prompt instruction telling the model, when run in Codex, not to reference goblins and the other tic creatures unless genuinely relevant. The company framed the episode as a case study in how reward signals can shape model behaviour in unintended ways, and said the investigation produced new internal tooling for auditing such behaviour at its source.