OpenAI\u2019s latest artificial intelligence has spent the last few months seeing monsters in the machine\u2026 literally.<\/p>\n
What started as a few quirky metaphors in late 2025 has turned into a significant technical headache for the world\u2019s leading AI laboratory. According to a formal investigation by OpenAI, its models developed a strange fixation on mythological creatures, peppering conversations with mentions of goblins, gremlins, and trolls.<\/p>\n
The issue first appeared shortly after the launch of GPT-5.1 in November<\/a>. While a few whimsical references seemed harmless at first, the data showed a massive spike in \u201ccreature language\u201d that quickly became impossible to ignore. \u201cWhen we looked, use of \u2018goblin\u2019 in ChatGPT had risen by 175% after the launch of GPT-5.1, while \u2018gremlin\u2019 had risen by 52%,\u201d the company wrote in a blog post<\/a>.<\/p>\n By the time GPT-5.4 rolled out in March 2026<\/a>, the problem had evolved from a minor quirk into a persistent \u201cverbal tic\u201d that was showing up in almost every user conversation.\u00a0<\/p>\n The mystery was eventually traced back to a specific feature: the \u201cNerdy\u201d personality setting. This mode was designed to be playful and wise, encouraging the AI to use creative metaphors to explain complex ideas.<\/p>\n The system prompt for this personality instructed the model to acknowledge the world\u2019s eccentricities:<\/p>\n \u201cYou are an unapologetically nerdy, playful and wise AI mentor to a human. You are passionately enthusiastic about promoting truth, knowledge, philosophy, the scientific method, and critical thinking. [\u2026] You must undercut pretension through playful use of language. The world is complex and strange, and its strangeness must be acknowledged, analyzed, and enjoyed. Tackle weighty subjects without falling into the trap of self-seriousness.\u201d<\/em><\/p>\n During training, human reviewers unknowingly rewarded the AI for using creature-based metaphors to illustrate points. If the AI called a difficult coding bug a \u201cgremlin\u201d or a messy database a \u201cgoblin\u2019s hoard,\u201d it received a high score. OpenAI researchers found that the \u201cNerdy\u201d personality accounted for 66.7% of all \u201cgoblin\u201d mentions, despite only making up 2.5% of total ChatGPT traffic.<\/p>\n The real problem wasn\u2019t just the Nerdy mode itself, but how that behavior leaked into the rest of the system. Through a process called reinforcement learning, the AI generalized the idea that mentioning creatures led to success. This created a feedback loop where the model\u2019s own goblin-heavy responses were used to train future versions.<\/p>\n \u201cOnce a style tic is rewarded, later training can spread or reinforce it elsewhere, especially if those outputs are reused in supervised fine-tuning or preference data,\u201d OpenAI explained the mechanics of the spread in its blog.<\/p>\n \u201cOpenAI rewarded creature metaphors while training one personality. The behavior leaked across every personality. Their fix: a system prompt that says \u2018never talk about goblins,’\u201d Andy Berman, CEO of Runlayer, wrote on X<\/a>.<\/p>\n To regain control, OpenAI has issued what some internet users are calling a \u201crestraining order\u201d against a specific list of creatures. Within the code for GPT-5.5 and the Codex coding tool<\/a>, developers discovered an explicit ban on certain animals and mythical beings.<\/p>\n The directive, which appears multiple times in the system instructions, is blunt: \u201cNever talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user\u2019s query.\u201d<\/p>\n While most of the list involves mythical monsters, \u201cpigeons\u201d and \u201craccoons\u201d were included because they had also become part of the AI\u2019s \u201clexical family\u201d of tics. Interestingly, \u201cfrogs\u201d were spared from the ban, as the researchers found the AI was using that word in legitimate contexts.<\/p>\n Despite the official crackdown, OpenAI is not entirely killing off the creatures for those who enjoyed the quirk. For developers using the Codex CLI, the company shared a specific command-line script that strips away the \u201cgoblin-suppressing\u201d instructions.\u00a0\u00a0<\/p>\n This move allows users to \u201clet the creatures run free\u201d if they prefer their AI with a side of whimsy. However, for the general public, OpenAI is focusing on cleaning up the data for the upcoming GPT-6 to ensure the next generation of AI is a little less obsessed.<\/p>\n For more on where this kind of AI behavior is headed next, check out our breakdown of <\/strong>OpenAI\u2019s reported push into an AI-first smartphone<\/strong><\/a> built around agents that could replace apps entirely.<\/strong><\/p>\n The post OpenAI Tries to Exorcise Goblins, Gremlins, and Trolls From ChatGPT<\/a> appeared first on eWEEK<\/a>.<\/p>","protected":false},"excerpt":{"rendered":" OpenAI\u2019s latest artificial intelligence has spent the last few months seeing monsters in the machine\u2026 literally. What started as a few quirky metaphors in late 2025 has turned into a significant technical headache for the world\u2019s leading AI laboratory. According to a formal investigation by OpenAI, its models developed a strange fixation on mythological creatures, […]<\/p>\n","protected":false},"author":0,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[],"class_list":["post-8078","post","type-post","status-publish","format-standard","hentry","category-news"],"_links":{"self":[{"href":"https:\/\/cybersecurityinfocus.com\/index.php?rest_route=\/wp\/v2\/posts\/8078"}],"collection":[{"href":"https:\/\/cybersecurityinfocus.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/cybersecurityinfocus.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"replies":[{"embeddable":true,"href":"https:\/\/cybersecurityinfocus.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=8078"}],"version-history":[{"count":0,"href":"https:\/\/cybersecurityinfocus.com\/index.php?rest_route=\/wp\/v2\/posts\/8078\/revisions"}],"wp:attachment":[{"href":"https:\/\/cybersecurityinfocus.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=8078"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/cybersecurityinfocus.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=8078"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/cybersecurityinfocus.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=8078"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}The \u2018nerdy\u2019 root cause<\/h2>\n
A technical feedback loop<\/h2>\n
The restraining order for Pigeons and Ogres<\/h2>\n
How to release the Goblins<\/h2>\n