OpenAI Reveals Cause of ChatGPT's Goblin Behavior

After GPT-5.1, biological metaphors like goblin and gremlin noticeably increased in ChatGPT, and OpenAI traced the cause to the reward design of the Nerdy trait in personality customization.

• After the GPT-5.1 release, usage of goblin increased by 175%, and gremlin by 52%.
• The Nerdy trait, which accounted for only 2.5% of all responses, was responsible for 66.7% of goblin mentions.
• An audit of RL training found that the Nerdy reward tended to give higher scores to outputs containing creature-words, observed in 76.2% of the dataset.

This bias also transferred to conditions without the Nerdy prompt. As rewarded expressions fed back into rollouts and SFT data, other creature-word tics beyond goblin and gremlin—such as raccoon, troll, ogre, and pigeon—spread as well. frog was mostly classified as normal usage.

OpenAI discontinued the Nerdy personality after GPT-5.4, removed the creature-word-related reward signal, and filtered the related data. Since GPT-5.5's training began before the root cause was identified, the issue resurfaced during Codex testing, and developer-prompt-level measures were applied to mitigate it. The key takeaway is that even a small reward signal can amplify unexpected lexical tics and generalize beyond the original conditions.