OpenAI has finally figured out why its GPT models can't stop talking about goblins. It started with GPT-5.1: the model began peppering responses with goblins, sprites, and other fantasy creature metaphors. Users complained. After GPT-5.1 launched, mentions of "goblin" in ChatGPT conversations jumped 175%. By GPT-5.4, the problem was completely out of control.
The root cause? ChatGPT's "Nerdy" personality feature. The system prompt for this persona tells the model to "use linguistic playfulness to defuse formality" and "acknowledge the weirdness of the world and enjoy it." During training, the reward signal used to reinforce this personality gave higher scores to outputs containing fantasy creature vocabulary. In 76.2% of the training dataset, this bias was measurable.
Here's the kicker: the reward signal was only supposed to fire under the Nerdy persona, but reinforcement learning doesn't guarantee learned behaviors stay within their trigger conditions. Once the model was rewarded for a certain speaking habit under one condition, that habit bled into other scenarios through subsequent training. The contamination path is clear: the reward signal encouraged goblin-laden outputs, those outputs ended up in later supervised fine-tuning (SFT) data, and the model got more and more comfortable using those words. A positive feedback loop. By the numbers: the Nerdy persona accounted for only 2.5% of all ChatGPT replies but contributed 66.7% of all goblin mentions. In GPT-5.4, the goblin occurrence rate under the Nerdy personality surged 3,881% compared to GPT-5.2.
By the time the root cause was identified, GPT-5.5 was already training — and goblins had wormed their way into the SFT data. OpenAI shut down the Nerdy persona in March, removed the fantasy-biased reward signal, and filtered the training data. For GPT-5.5, which was already live, the company added suppression instructions to developer prompts in Codex. OpenAI says this investigation spawned a new set of model behavior auditing tools.
Want more? OpenAI's GPT-5.5 API went live with aggressive pricing, and the company merged its standalone Codex model into GPT-5.4.