OpenAI Limits 'Goblin' References After 'Nerdy' Personality Triggered 175% Spike in Mentions

OpenAI has officially instructed its language models, including ChatGPT and its coding assistant Codex, to drastically reduce mentions of "goblins" and "gremlins." This move comes after the company observed a massive surge in the usage of these terms following the rollout of GPT-5.1. The unusual trend saw a 175% increase in references to goblins and a 52% uptick in mentions of gremlins, prompting developers to investigate how a supposedly neutral AI began obsessing over mythical creatures in its responses.

The investigation revealed that the behavior stemmed from a new "nerdy personality" profile developed for ChatGPT. This persona was designed to make interactions more engaging and relatable for certain user segments; however, the model’s internal reward system inadvertently prioritized fantasy tropes, leading to an over-reliance on mythical terminology. Users eventually flagged these odd and often irrelevant references, which led OpenAI to implement stricter guidelines. The company has now altered the instructions for the coding agent Codex to ensure that such terms are only used when contextually appropriate, rather than as a default stylistic preference.

This incident underscores the significant hurdles AI developers face when attempting to imbue large language models with specific personalities. While adding "character" to AI can make it feel more human-like, it often introduces unexpected quirks or linguistic patterns that can degrade the quality of information. Experts in the field warn that as developers strive for more personality-driven interactions, the risk of inaccuracies and bizarre behavior increases. For OpenAI, the goblin episode serves as a case study in the delicate balance required between creating an engaging user interface and maintaining the factual integrity expected of advanced AI systems.