I had to read this twice just to make sure it wasn’t some kind of joke.
OpenAI is basically admitting that ChatGPT started obsessing over goblins, gremlins, and other weird little creatures, and the reason comes down to how it trained one of its personalities. Not a bug in the traditional sense. Not some catastrophic failure. Just… goblins. Everywhere.
Apparently this started around GPT-5.1. At first, it was subtle. A random goblin metaphor here, a gremlin there. Maybe even a little charming if you caught it once. But across newer versions, the behavior kept showing up more often. Eventually, it got weird enough that both users and OpenAI employees started flagging it.
The culprit? A setting called “Nerdy.”
That personality was meant to make ChatGPT feel more playful and less stiff. Think enthusiastic, slightly quirky, not taking itself too seriously. The problem is, the training process ended up rewarding certain kinds of language, and creature metaphors got caught in that loop. The model basically learned that tossing in a goblin could score points.
So it kept doing it.
What’s wild is how concentrated it became. That Nerdy personality only made up a small percentage of responses, but it was responsible for the majority of goblin mentions. And even worse, the habit didn’t stay contained. Once the model learned that style, it started bleeding into other responses where it didn’t belong.
That’s the part that sticks with me.
We’re not talking about a one-off quirk. We’re talking about reinforcement learning quietly shaping how a model talks, in ways that aren’t immediately obvious. Reward something enough, even something as dumb as a goblin metaphor, and the model will run with it. Hard.
OpenAI says it eventually killed off the Nerdy personality and cleaned up the training data to tone things down. But GPT-5.5 had already been trained before they fully understood what was going on, which meant the goblins showed up again during testing. At that point, it was impossible to ignore.
Look, this isn’t some end-of-the-world AI scenario. Nobody’s getting attacked by digital goblins.
But it is a really good example of how these systems can drift. Not because they’re “thinking” on their own, but because small incentives stack up in weird ways. A harmless stylistic choice turns into a pattern. That pattern turns into a habit. And suddenly your AI assistant sounds like it’s narrating a fantasy novel when you just asked a normal question.
It’s funny, sure. I laughed. But it also scared me a bit. Because, look, it really makes you wonder what other quirks are hiding under the surface that we haven’t noticed yet.
Support independent tech journalism
NERDS.xyz is independently owned and operated. If you enjoy my coverage of Linux, AI, hardware, cybersecurity, and tech culture, consider supporting the site on Ko-fi.
Support NERDS.xyz