OpenAI’s Goblins, Gremlins, and the Strange Habits of AI Models

OpenAI has a goblin problem. Actually, it’s more of a goblin, gremlin, raccoon, troll, ogre, and pigeon problem — a whole menagerie of mythical and real creatures that its coding model started obsessing over. After Wired reported that OpenAI had to explicitly instruct its model to “never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures,” the company finally broke its silence with a blog post explaining the weirdness.

The short version: it’s a side effect of training data and personality options. The longer version is more interesting, and it reveals how fragile these models can be.

OpenAI says the trouble started with GPT-5.1, specifically when users selected the “Nerdy” personality option. The model began dropping metaphors involving goblins and other creatures — not in a helpful way, but in a way that felt like a broken record. The problem got worse with each subsequent model update. The company calls it a “strange habit,” which is a polite way of saying the model developed a bizarre fixation that required manual intervention.

I’ve seen this kind of thing before. It’s not unlike how older language models would veer into talking about unicorns or dragons if the training data had too many fantasy novels. But OpenAI’s case is more specific: the creatures in question are often used in engineering and programming metaphors (goblins for messy code, trolls for bad actors, etc.), so the model probably latched onto them as a crutch. The “Nerdy” personality likely amplified that, turning a useful metaphor into a repetitive tic.

What’s frustrating is that OpenAI didn’t catch this earlier. The Wired report made it sound like the company had to hardcode a block list, which is a clumsy fix for a deeper issue. In the blog post, OpenAI says it’s working on better training techniques to prevent these “habits” from forming. But honestly, this feels like a band-aid. The real fix is better data curation and more robust testing for personality modes — not just telling the model to shut up about goblins.

Still, I’ll give OpenAI credit for being transparent about it. Most companies would bury this in a footnote or ignore it entirely. The blog post is short and doesn’t sugarcoat the problem. It’s a reminder that even the most advanced AI models are still prone to strange, unpredictable behaviors. And if you’re using the “Nerdy” option, you might want to keep an eye out for raccoon-related tangents.