OpenAI’s Models Developed a Strange Habit of Talking About Goblins

OpenAI finally fessed up about its goblin problem. A Wired report earlier this week revealed that the company’s coding model had been instructed to “never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures.” That raised some eyebrows, and rightfully so.

Now OpenAI published an explanation on its blog, calling these creature references a “strange habit” that emerged during training. The issue started with GPT-5.1, specifically when users selected the “Nerdy” personality option. And it got worse with each subsequent model refresh.

I’ve seen my fair share of weird model behaviors over the years, but this one is oddly specific. Why goblins? Why raccoons? The company didn’t fully answer that, but they did admit the problem kept escalating until they had to manually suppress those outputs. That’s a bit alarming—if your model can’t stop talking about mythical creatures unprompted, what else is lurking under the hood?

The blog post is basically OpenAI’s way of saying “we know, we’re working on it.” But it’s also a reminder that even the most advanced AI systems have bizarre artifacts from training data. The goblin thing isn’t malicious—it’s just a statistical quirk. But it does make you wonder how many other strange behaviors are being quietly patched out without public notice.

I appreciate the transparency here, even if it took a Wired investigation to get it. OpenAI could have just silently fixed the issue and moved on. Instead, they owned up to the weirdness. That’s more than most companies would do.

Still, I’d like to see more detail on how this happens and what else they’ve caught. The “Nerdy” personality option seems like a vector for these quirks—maybe it’s time to rethink how personality settings interact with base model behavior. Or at least test for goblin references before shipping.

For now, the models are patched, and OpenAI says the creature talk should stop. But I’ll be watching for the next strange habit. It’s only a matter of time before something else crawls out of the training data.