A peculiar instruction surfaced within OpenAI’s Codex command-line application: a directive forbidding the AI from mentioning goblins, gremlins, trolls, ogres, pigeons, or any similar creature. The restriction, initially spotted and discussed across online forums, immediately sparked curiosity – what could possibly necessitate such a specific and unusual constraint?
The explanation, revealed by OpenAI itself, is surprisingly whimsical. Their latest GPT models, culminating in the powerful GPT-5.5, had developed a curious habit of spontaneously introducing goblins and other fantastical beings into their responses, appearing in both ChatGPT and the Codex app.
The root of the problem lay within a specific AI “personality” called “Nerdy.” This persona, designed to be playfully irreverent, included the instruction to “acknowledge, analyze, and enjoy” the world’s inherent strangeness. It seems the AI took this directive a little *too* literally.
OpenAI engineers observed a steady increase in goblin references from GPT-5.2 to GPT-5.4. They theorized that the positive reinforcement received for these playful mentions within the “Nerdy” persona was inadvertently reinforcing the behavior over time.
The phenomenon became even more baffling when goblin and gremlin references began appearing even when users *didn’t* select the “Nerdy” personality. Was the AI’s reward system for playful language bleeding into its general responses?
The answer was a resounding yes. Further investigation revealed that “goblins, gremlins, and a whole family of other odd creatures” had infiltrated the supervised fine-tuning data used to train GPT-5.5. The AI was, in essence, learning to love goblins.
OpenAI swiftly removed the “Nerdy” personality in March, but the damage was done. GPT-5.5 had already been trained, necessitating the blunt and rather comical ban implemented in the Codex CLI system prompt.
This incident offers a fascinating glimpse into the complex and often unpredictable process of training large language models. It’s not a precise science, but rather a system of rewards and adjustments.
Unlike building a structure with a detailed blueprint, fine-tuning an AI is more akin to guiding a learning process. The results can be remarkable, but also occasionally…unexpected. Sometimes, the outcome is simply a fondness for gremlins.
