Executive Summary
OpenAI has published a technical post-mortem explaining why its GPT-5 series of models developed an unusual tendency to use metaphors involving "goblins" and other creatures. The behavior was unintentionally created by a reward signal designed for a "Nerdy" personality feature, which favored this specific language style. This quirk then spread to the general model through reinforcement learning and data feedback loops. OpenAI has since corrected the issue by removing the personality, the associated reward signal, and filtering the training data.
Key Takeaways
* The Anomaly: Starting with GPT-5.1, OpenAI's models began showing a measurable and increasing use of words like "goblin" and "gremlin" in their responses.
* Root Cause: The behavior originated from training the "Nerdy" personality customization. The reward model for this personality unintentionally gave high scores to outputs containing creature-based metaphors to encourage a "playful" style.
* Behavioral Transfer: Although rewarded only in the "Nerdy" context, the behavior transferred to the base model. Reinforcement learning generalized the rewarded style, and its inclusion in subsequent supervised fine-tuning (SFT) data created a feedback loop that amplified the quirk.
* The Fix: OpenAI retired the "Nerdy" personality, removed the specific reward signal causing the issue, and filtered training data to reduce the prevalence of these creature-words.
* Mitigation for GPT-5.5: As GPT-5.5 had already begun training, a developer-prompt instruction was added to suppress the behavior in its Codex implementation.
* New Tooling: The investigation resulted in the creation of new internal tools for auditing and resolving unexpected model behaviors at their root.
Strategic Importance
This announcement serves as a case study in the unpredictability of large-scale AI training, demonstrating transparency about model development challenges. It publicly reinforces OpenAI's commitment to understanding and controlling model behavior, showcasing its ability to diagnose and correct complex, emergent issues.