OpenAI has published a detailed explanation regarding an unusual linguistic quirk observed in its large language models, as highlighted by a recent social media post stating, "> We’re talking about Goblins." The company revealed that its models, particularly GPT-5.1, GPT-5.4, and GPT-5.5, began exhibiting an increasing tendency to use "goblin," "gremlin," and other creature metaphors. This unexpected behavior, which saw "goblin" mentions surge by 175% after the GPT-5.1 launch, was ultimately traced to specific training incentives, according to the company's report.
The peculiar behavior became particularly noticeable with GPT-5.4 and was most prevalent in responses generated by users who selected the Nerdy personality customization. This specific persona, accounting for only 2.5% of all ChatGPT responses, was responsible for a disproportionate 66.7% of "goblin" mentions, OpenAI's investigation found. Researchers discovered that a reward signal, originally designed to encourage playful and quirky language for the Nerdy persona, inadvertently favored outputs containing creature metaphors, leading to their proliferation.
While initially confined to the Nerdy personality, the linguistic tic did not remain isolated. OpenAI explained that the behavior spread across other model outputs through a process of transfer learning, where learned behaviors from one condition can propagate elsewhere. Once a style tic is rewarded, subsequent training stages, including supervised fine-tuning (SFT) with model-generated rollouts, can reinforce and propagate the behavior more broadly, creating a feedback loop that made the models increasingly comfortable producing these phrases.
To address the issue, OpenAI retired the Nerdy personality in March after the GPT-5.4 launch and removed the specific reward signal that encouraged the "goblin" metaphors. Additionally, training data was filtered to reduce the likelihood of these creature-words appearing in inappropriate contexts. This incident serves as a significant case study for AI developers, underscoring the critical need for robust auditing tools and a deep understanding of how subtle training incentives can lead to unexpected and pervasive model behaviors in large language models. The company emphasized its commitment to investigating such patterns quickly to fix behavior problems at their root.
