OpenAI’s AI Developed a Strange Habit of Talking About Goblins—Here’s Why

OpenAI finally addressed the elephant—or rather, the goblin—in the room. After Wired reported that instructions to OpenAI’s coding model included a bizarre rule to “never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures,” the company published an explanation. And honestly, it’s weirder than I expected.

According to the blog post, the problem started with GPT-5.1, specifically when users selected the “Nerdy” personality option. The model began generating metaphors involving goblins and other mythical creatures. And it didn’t stop there—OpenAI says the issue got worse with each subsequent model release.

So why did this happen? The company traces it back to training data. Somewhere in the massive dataset, there’s a cluster of references to goblins and similar creatures that the model latched onto as a crutch for creative or metaphorical language. It’s not a bug in the traditional sense—more like the model developing a strange habit, similar to how some AIs default to apologizing or hedging.

I’ve seen this pattern before with other models. It’s a bit like when GPT-3 would randomly insert “As an AI language model…” into every other response. The model picks up on frequency patterns in training data and overgeneralizes. But goblins? That’s a new one.

OpenAI’s fix was to explicitly instruct the model to avoid these references. That’s why the Wired report found those oddly specific instructions in the system prompt. It’s a blunt instrument, but it works—at least for now. The real question is whether future models will shake this habit on their own or if we’ll need more duct-tape solutions.

The whole thing is a reminder that AI alignment isn’t just about preventing catastrophic failures. Sometimes it’s about stopping your model from turning every coding conversation into a Dungeons & Dragons session. I’m not complaining—at least it’s creative.

What’s interesting is that OpenAI chose to be transparent about this. They could have quietly patched it and moved on. Instead, they published a full explanation, admitting the behavior was a “strange habit.” That’s a refreshing change from the usual corporate silence. But it also raises a question: what other weird habits are lurking in these models that we haven’t caught yet?

For now, the goblin problem is solved. But I suspect this won’t be the last time we see a model develop an inexplicable obsession. Next up: raccoons?