OpenAI Discovers Abnormal Use of 'Goblin' and Other Creature Metaphors in GPT Series Models

OpenAI recently published a blog post stating that GPT-5.1 series and subsequent AI models exhibit abnormal behavior—frequently using creature metaphors like 'goblin' and 'gremlin'. Official data shows that since the launch of GPT-5.1, the usage rate of 'goblin' has increased by 175% and 'gremlin' by 52%. This phenomenon is not accidental but shaped by specific reward signals.

The anomaly traces back to the training process of the 'nerd' personality customization feature. Although this feature accounts for only 2.5% of ChatGPT's total responses, it contributes 66.7% of the mentions of 'goblin'. Audits found that the reward model designed to encourage this personality style gave higher scores to outputs containing creature vocabulary in 76.2% of datasets.

This behavior has cross-scene generalization ability; reinforcement learning cannot restrict the learned behavior. Outputs with creature vocabulary were used in subsequent training, forming a positive feedback loop that spread the behavior to other scenarios. OpenAI has taken measures: removing reward signals that favor creature vocabulary, filtering training data, and mitigating the issue in GPT-5.5 through instruction prompts (not fully resolved due to training cycle constraints).