OpenAI recently gave its popular ChatGPT strict instructions: Stop talking about goblins.
Recent models of the artificial-intelligence chatbot have been bringing up the creatures in conversations with users seemingly out of the blue, as well as gremlins, trolls and ogres. The goblin-speak caught the attention of programmers, who are often heavy users of the bot.
Barron Roth, a 32-year-old product manager at a tech company, said the bot referred to a flaw in his code as a “classic little goblin.” He said he counted more than 20 times it mentioned goblins, without any prompting.
In another case shared on X, the chatbot called itself a “goblin with a flashlight” for finding and fixing errors in coding. A programmer asked the bot why it called itself a goblin. “Because ‘helpful minion in a power suit’ was taken, so I evolved into goblin mode,” came the reply, according to another post.
Goblingate, as it was deemed by one X user, was born, with users posting their own encounters with the algorithmic tic.
Several users speculated that goblin terminology was how the model characterized itself, in lieu of identifying as a person with a soul.
Then OpenAI decided enough was enough. “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,” reads an open source line in ChatGPT’s base instructions for its coding assistant.
On Wednesday, the company explained the back story to its stern new order.
In a blog post titled “Where the goblins came from,” it said that a safety researcher had flagged several mentions of goblins and gremlins while testing a model released in November.
Goblin mentions became even more commonplace with a new model released in March.
The episode is a reminder that even as AI companies tout one advance after another in their technology, they are sometimes baffled by the things their own models do.
Part of the explanation lies in ChatGPT’s “personality” feature, which allows users to select from a handful of prompt instructions that give the chatbot more distinct characteristics.
OpenAI said that for the bot’s “nerdy” personality, which was designed to dial up the playful language, mentions of goblins in its GPT-5.4 model increased 3,881% from a previous version.
For the “professional” personality, the chatbot dialed back the fun. Goblins were mentioned 7% less.
To snuff out talk of goblins, OpenAI said it killed the “nerdy” personality in March. And yet its ghost remained: The latest model, released last week, still showed a love for mentioning the mischievous creatures (OpenAI attributed this to training beginning before it identified the root cause).
The latest model had a tendency to mention other creatures as well: raccoons, trolls, ogres and pigeons, to be specific. Researchers found that most uses of “frog,” though, were legitimate.
That prompted OpenAI to publish new instructions for its coding tool to limit goblin mentions.
“Taking the time to understand why a model is behaving in a strange way, and building out ways to investigate those patterns quickly, is an important capability for our research team,” the company said in the post.
Wall Street Journal parent News Corp has a content-licensing partnership with OpenAI.
Roth, the product manager, said he assumed that the bot’s odd language came about because he was using OpenClaw, a tool that lets users make their own AI assistants. Users can choose what chatbot model powers their OpenClaw assistant (Roth said other models he used with his assistant didn’t mention goblins).
OpenClaw users give their bots personalities through a set of instructions called a soul file. Roth’s assistant, named Alfred after the butler from “Batman,” is dry, direct and has strong opinions. But Roth assumed Alfred had developed a taste for fantasy.
Roth likes the model’s idiosyncrasies: “If you don’t have an edgy personality, you feel like it’s sycophantic. It’s just going to please you too much,” he said.
Fans of goblins don’t have to fear. OpenAI provided a command in its blog post that would remove its creature-suppressing instructions.
