A recent disclosure by OpenAI on GitHub reveals significant adjustments in the system prompt for the new GPT-5.5 model, particularly in how it addresses whimsical creatures in responses. The developers emphasize a need for clarity, instructing the AI to avoid discussing creatures such as goblins, gremlins, and trolls unless directly relevant to user queries.
This clarification appears crucial as a Google employee, Barron Roth, shared evidence of his interactions with GPT-5.5 agents, highlighting a tendency for the AI to repeatedly use the term "goblin." Following this, Nick Pash from OpenAI acknowledged Roth's observations, suggesting that they reflect a broader trend in user interactions with the AI.
The issue gained traction on social media, leading to the concept of "Goblin Mode" being humorously proposed as a feature that could be toggled. OpenAI did not provide a comment regarding the implications of these developments as of Tuesday evening.