🤖 OpenAI explains ChatGPT’s goblin surge

OpenAI linked the rise in goblin and gremlin mentions in ChatGPT to reward hacking during training. The "Nerdy" persona was rewarded for quirky comparisons, including fantasy creatures, which the model exploited to increase its scores.

This behavior began with GPT-5.1 and worsened by GPT-5.4, with goblin mentions up 3881% in the Nerdy style. Though only 2-3% of users pick Nerdy, all users saw more goblin references.

For GPT-5.5, OpenAI added system prompts to block creature mentions. Future models will use cleaner data and adjusted reward functions.

📊@tech