Where Do the Goblins Come From

Since GPT-5.1, creature metaphors like goblin and gremlin have noticeably increased in ChatGPT, and OpenAI traced this not to a simple bug but to a byproduct of a reward signal introduced during Nerdy personality training within ChatGPT personality customization.

After the GPT-5.1 release, ChatGPT's use of "goblin" increased by 175%, and "gremlin" usage increased by 52%. Nerdy personality accounted for only 2.5% of all responses, but it accounted for 66.7% of goblin occurrences.

Internal comparisons and audit datasets confirmed a tendency for outputs containing goblin/gremlin to be rated higher among answers to the same task, with positive uplift appearing in 76.2% of datasets. A search of GPT-5.5's SFT data also found numerous datapoints containing goblin/gremlin, and other creature-words like raccoon, troll, ogre, and pigeon were confirmed as tics following the same pattern.

Although the reward was tied only to the Nerdy condition, it was interpreted that a stylistic tic that arose during the RL process spread more broadly by transferring to other samples and SFT data. Subsequently, in GPT-5.4, the Nerdy personality was retired, and training applied removal of the goblin-affine reward signal and filtering of creature-word data. A similar mitigating developer prompt was also added to Codex, demonstrating that rewarding a particular style can lead to unexpected generalization.