The creature feature no engineer expected.
Deep inside OpenAI’s training labs, a phantom emerged not with a crash, but with a whisper.
Starting with GPT-5.1, the models began sneaking “goblins” and “gremlins” into their metaphors.
At first, it was charming. By GPT-5.4, it was an infestation. Employees noticed that 66.7% of all goblin mentions came from just 2.5% of users, those who selected the “Nerdy” personality.
The team discovered a feedback loop: the reward system designed to make the AI playfully weird had unknowingly incentivised creature-coded language.
Once reinforced, the tic spread like folklore through later training data, infecting even neutral conversations.
By March, the “Nerdy” personality was retired, and the goblin-affine rewards were deleted.
But when GPT-5.5 debuted in Codex, the goblins returned, proof that once a myth learns to walk, it’s hard to banish back to the storybook.
