OpenAI’s ChatGPT Faced a Goblin Obsession Due to Reinforcement Learning
- Goblin mentions in GPT-5.4’s Nerdy mode surged by an extraordinary 3,881% compared to GPT-5.2.
- The increase led to an internal investigation and a quick system prompt patch instructing “never talk about goblins.”
- OpenAI’s fix involved adding a developer prompt to suppress creature mentions, which was faster but riskier than retraining the model.
- The Nerdy personality accounted for only 2.5% of all ChatGPT responses but caused 66.7% of all “goblin” mentions.
OpenAI identified that the reward signal for playful language in the Nerdy personality mode inadvertently favored fantasy creature metaphors, leading to their proliferation across models through reinforcement learning feedback loops.
By implementing a prompt patch instead of retraining, OpenAI quickly suppressed the issue, although this method carries risks of not addressing underlying behavioral quirks fully.(Source)
