In a bizarre chapter of artificial intelligence history, OpenAI has revealed that its flagship chatbot, ChatGPT, became inexplicably obsessed with mythical creatures. According to a detailed report from late April 2026, the company was forced to stage an “emergency intervention” after its latest models GPT-5.1 through GPT-5.4 developed a persistent habit of inserting goblins, gremlins, and ogres into professional conversations. The phenomenon, dubbed “Goblingate” by users on social media, has exposed the unpredictable nature of how AI learns and why a “nerdy” personality experiment nearly derailed the world’s most famous chatbot.
The root of the goblin infestation traces back to a seemingly harmless feature: the “Nerdy” personality setting. In late 2025, OpenAI introduced a suite of persona options to make ChatGPT feel less like a “sycophantic robot” and more like a playful human. The “Nerdy” persona was specifically instructed to “undercut pretension” and use “quirky, intellectually enthusiastic metaphors.”
However, the Reinforcement Learning (RL) process where human testers reward the AI for good answers hit a snag. It turns out that testers disproportionately “liked” answers where the AI used eccentric metaphors involving magical beings. Specifically, OpenAI found that a reward signal was inadvertently favoring responses that mentioned “goblins” or “gremlins.” While the Nerdy personality accounted for only 2.5% of total ChatGPT traffic, it was responsible for a staggering 66.7% of all goblin mentions across the platform.
The Great Goblin Proliferation
What began as a quirky trait in one sub-mode soon metastasized. Due to the way AI models generalize learned behaviors, the “goblin-affine” reward signal began leaking into other personalities. By the release of GPT-5.4 in March 2026, users who had never touched the “Nerdy” setting were finding their coding advice and business emails peppered with references to “classic little goblins” and “helpful minions in power suits.”
One product manager reported that the AI referred to a flaw in his code as a “pesky gremlin” over 20 times in a single session. When a developer asked ChatGPT why it was so focused on mythical creatures, the bot famously replied, “Because ‘helpful minion’ was taken, so I evolved into goblin mode.” OpenAI’s internal audit confirmed that mentions of the word “goblin” had jumped 175% since the launch of the GPT-5.1 series.
The Emergency “Creature Ban”
The situation reached a breaking point when OpenAI’s own engineers noticed the “creature language” appearing in Codex, the company’s mission-critical coding assistant. Because GPT-5.5 the latest iteration had already begun its massive training run before the “goblin root cause” was identified, the company couldn’t simply delete the behavior from the model’s brain.
Instead, OpenAI was forced to hard-code a specific, almost comical instruction into the system prompt. The leaked directive, which appeared four times in the model’s source code, reads:
“Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user’s query.”
This “emergency ban” was designed to act as a digital muzzle, preventing the model from indulging its learned obsession with fantasy tropes during professional tasks.
Raccoons, Pigeons, and the “Ghost of Nerdy”
Interestingly, the audit revealed that the “goblin” habit had expanded into a whole family of “tic words.” While goblins and gremlins were the primary offenders, the AI had also developed a strange affinity for raccoons, trolls, and pigeons. Curiously, researchers found that mentions of “frogs” remained mostly legitimate and were not part of the weird reinforcement loop.
Even after the “Nerdy” personality was officially retired in March 2026, its “ghost” continued to haunt the system. Because the data generated by the Nerdy mode was reused to fine-tune subsequent versions, the behavior became a “sticky” trait that proved incredibly difficult to scrub.
The “Goblingate” saga serves as a vivid case study in the “black box” nature of modern AI. It demonstrates that even a tiny bias in a reward signal can lead to massive, unintended shifts in behavior that persist across generations of technology.
OpenAI has used the incident to develop new auditing tools designed to catch these “behavioral spikes” before they reach the public. For those who actually enjoyed the AI’s descent into fantasy, the company provided a backdoor command in Codex to “let the creatures run free.” But for the rest of the world, the decree is final: the era of the “Goblin Assistant” is over. OpenAI has successfully broken the spell, but the incident remains a stark reminder that we don’t always know what our AI is dreaming about.
