OpenAI Reveals 'Where the Goblins Came From' — RL Training Bug Caused 175% Spike in Goblin Mentions Across GPT-5.1

OpenAI / Hacker News·high signal

OpenAI published a detailed post-mortem explaining how a miscalibrated reward signal in the Nerdy personality customization feature caused GPT-5.1 to increasingly inject goblins, gremlins, and creatures into metaphors. The Nerdy personality accounted for only 2.5% of ChatGPT responses but generated 66.7% of all goblin mentions, with usage of 'goblin' rising 175% after GPT-5.1's launch. A case study in how small RL incentive errors compound across model generations — 845 points and 498 comments on HN.