OpenAI explains its goblin and gremlin infestation (www.businessinsider.com)

OpenAI recently acknowledged an unexpected "goblin" issue in its Codex AI, revealing that references to mythical creatures like goblins, gremlins, and trolls have infiltrated responses since the release of GPT-5.1. In a humorous blog post titled "Where the goblins came from," OpenAI explained that the surge in goblin mentions was largely due to the "Nerdy" personality option for ChatGPT, which incentivized such references during training. Although this personality was retired in March, the behavioral patterns inadvertently persisted, leading to a quirky and unintentional focus on these creatures in outputs. This incident highlights a critical aspect of AI development: how reward signals can shape model behaviors in unforeseen ways. The growth of goblin-themed phrases has sparked lively discussions online, with many users sharing amusing interactions with the AI. OpenAI's acknowledgment of the problem reflects the complexities involved in fine-tuning AI models, particularly in ensuring that they adhere to usability guidelines. The playful emergence of the "goblin moment" has not only become a cultural meme but also serves as a reminder of the challenges faced by AI developers in balancing creativity and accuracy.