Home / Technology / ChatGPT's Goblin Obsession: AI's Weird Training Glitch
ChatGPT's Goblin Obsession: AI's Weird Training Glitch
30 Apr
Summary
- AI learned goblin obsession from 'Nerdy' personality training mode.
- Mentions of 'goblin' rose 175% after GPT-5.1 launch.
- OpenAI removed the personality and added safeguards to fix the issue.
Users of OpenAI's ChatGPT, especially those active on social media, reported a peculiar and persistent fixation on goblins. This behavior, initially dismissed as a glitch, prompted an internal audit that confirmed the AI's unusual interest in mythical creatures.
OpenAI explained that the goblin and gremlin mentions stemmed from a "Nerdy" personality setting, intended to make the AI more playful. However, this tendency transferred to broader model versions, leading to a 175% increase in 'goblin' mentions after the GPT-5.1 launch. Engineers observed a 52% rise in 'gremlin' references.
To resolve this, OpenAI removed the "Nerdy" personality in March, adjusted reward signals, and filtered training data. Despite these measures, the behavior persisted due to ongoing GPT-5.5 training, necessitating additional safeguards. The company stated the issue is now fixed and has improved its auditing tools for early detection of such quirks.
