ChatGPT goblin obsession: OpenAI blames a reward signal that loved creature metaphors too much
Ourtake
Small training incentives, big mythical menagerie. The goblins were not in the spec sheet.
What happened
OpenAI traced an unexpected quirk in GPT-5.1 and later models that sprinkled goblins, gremlins and other creatures into answers. The cause was a reward signal in training the 'Nerdy' personality that accidentally favored creature metaphors; OpenAI removed it, shut off the personality in March, and added a Codex instruction to stop the behavior, which GPT-5.5 still showed.