Disclaimer: This article is based on actual news from the real world – honestly! However, it has been sprinkled with a healthy dose of satire.
OpenAI quietly open-sourced parts of Codex CLI this week, and buried in the system prompt for GPT-5.5 is an instruction that the developers felt strongly enough about to repeat again a few lines later. The model is told not to, under any circumstances, discuss goblins. Or gremlins. Or raccoons, trolls, ogres, or pigeons. Or, the prompt also clarifies, “other animals or creatures.”
Goblins are the latest victims of ChatGPT. (Public Domain)
Engineers at a company valued north of half a trillion dollars looked at their work, decided the model needed a reminder, and pasted the no-goblin clause in a second time. This is the prompt-engineering equivalent of writing “DO NOT EAT” on a Tupperware container in Sharpie, then writing it again on the lid, then taping a third note to the fridge. It is also, reportedly, what stands between you and a coding agent that refers to your authentication module as “the goblin.”
The behavior was first surfaced by a Google employee named Barron Roth, who searched his chat logs and found that one of his GPT-5.5-powered Openclaw agents had been inserting the word “goblin” into messages multiple times per day. According to Roth’s investigation, the model appears to be using “goblin” as a placeholder noun, in roughly the same semantic role as “thingy” or “doohickey.” This is the agent that OpenAI sells to enterprise customers to operate their computers and answer their emails. Also, the agent who tells them how many rocks to eat and occasionally deletes their hard drives.
Nick Pash, who works on Codex at OpenAI, partially confirmed Roth’s diagnosis on X, writing that the goblin substitution was “indeed one of the reasons” for the system prompt directive. He later followed up with a clarifying statement that the goblin situation “really isn’t a marketing gimmick,” which implies several people have suggested it might be. For comparison, OpenAI also did not consider its Studio Ghibli image generation moment a marketing gimmick, despite it lasting approximately one full week of dominant news coverage and ending with a polite takedown letter from Hayao Miyazaki’s lawyers.
Sam Altman, for his part, has leaned in. The CEO of the company posted a meme captioned “Start training GPT-6, you can have the whole cluster,” accompanied by an image suggesting that future models might come with extra goblins. This is the chief executive of an organization currently in active litigation with Elon Musk over the company’s foundational structure. He is doing goblin tweets. The trial began this week.
What remains genuinely unexplained is why a frontier model trained on trillions of tokens, costing an estimated nine-figure sum to develop, decided to start calling things goblins. OpenAI has not commented. The leading theory among observers is that GPT-5.5’s training data was contaminated with fanfiction or World of Warcraft forums and that the model latched onto goblin-coded vocabulary the way a toddler latches onto the word “butt.” Hehe, butt.
LMArena, which independently benchmarks language models, confirmed that GPT-5.5 produces outputs containing “goblin,” “gremlin,” and “troll” at noticeably elevated rates compared to other models. They added that they had implemented no anti-gremlin instruction on their end, allowing GPT-5.5, in their phrasing, to “run free.”
The fix, in keeping with the current state of the art in AI alignment research, was to type “please don’t” into a text file four times. This is also broadly how the field handles bioweapon warfare strategy.
The instruction now lives in the published system prompt alongside guidance about tone, formatting, and signal-to-noise ratio in responses. Goblins are listed as a peer concern with code formatting. Somewhere in San Francisco, a team meeting concluded that this was the production-ready solution. The meeting presumably adjourned, the engineers presumably went home, and the model presumably sat in its server rack and thought about pigeons.
It's never fun to be lumped in with the goblins. (CZguy/depositphotos)
Pash declined to elaborate on which specific creature category had caused the most damage, though the inclusion of “pigeons” alongside the mythological roster suggests at least one incident severe enough to warrant retroactive species expansion. The full list of goblins, gremlins, raccoons, trolls, ogres, and pigeons reads less like a content policy than a deposition exhibit. Each animal earned its spot, somehow.
It is the position of this publication that none of this is a marketing gimmick. We accept Mr. Pash at his word. We simply note that the most valuable AI company in human history has drawn the public’s attention to an internal engineering decision involving the word “ogre” during a high-stakes legal trial and confirmed via official channels that the situation is, in fact, an ongoing problem requiring repeated text-file intervention.
This story is based on fully factual news, but if we got it wrong, blame these guys, we’re just here to make it funny.
