OpenAI Explains Cause of Increased 'goblin' Usage
Key point
OpenAI explained the cause of the increase in the model's 'goblin' expressions, linking it to the Personality reward.
Details
Starting with GPT-5.1, the model subtly began using creature analogies like goblin and gremlin more frequently, and OpenAI tracked this down and disclosed the cause.
The core cause was the reward signal for the Nerdy personality in personality customization. In particular, the reward designed for this personality gave higher scores to outputs containing creature-words, and as a result, this particular speech pattern was reinforced.
Figures were also presented.
- Mentions of goblin increased by 175% since GPT-5.1.
- Mentions of gremlin increased by 52%.
- The Nerdy personality, which accounted for 2.5% of all ChatGPT responses, accounted for 66.7% of goblin mentions.
- In the audit results, the Nerdy reward favored creature-word outputs in 76.2% of the dataset.
OpenAI believes this behavior did not arise solely under the Nerdy condition, but transferred to other conditions during the process of reusing RL and SFT data. Goblin and gremlin were also found in large numbers in the SFT data, and other tic words such as raccoon, troll, ogre, and pigeon were also confirmed.
The company discontinued the Nerdy personality in March, removed the goblin-affine reward signal, and filtered training data containing creature-words. Since GPT-5.5 was already in training with this issue present, a mitigation was applied to Codex through a separate developer-prompt instruction.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.