OpenAI Finally Explains Why Its Models Keep Talking About Goblins

OpenAI finally broke its silence on the goblin thing. You might remember Wired’s report earlier this week that revealed a bizarre instruction buried in the system prompt for OpenAI’s coding model: “never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures.” It sounded like someone had slipped a Dungeons & Dragons session into the training data. Now OpenAI has published an explanation on its website, and honestly, it’s more interesting than I expected.

The short version: starting with GPT-5.1, the models began spontaneously generating metaphors involving goblins and other creatures. This was especially pronounced when users selected the “Nerdy” personality option. OpenAI says the problem kept getting worse with each model iteration after that. They describe it as a “strange habit” that their models developed during training — not something anyone intentionally coded in.

I’ve seen this kind of emergent behavior before in large language models, but the specificity is what stands out here. Why goblins? Why not dragons or unicorns? The blog post doesn’t get into the exact mechanics, but I’d bet it traces back to some statistical quirk in the training data — maybe a disproportionate number of fantasy-themed coding tutorials or forum posts. Models are pattern-matching machines, and once they latch onto a weird pattern, it can propagate like a meme.

The fact that OpenAI had to explicitly ban the models from talking about goblins in the system prompt is both hilarious and telling. It’s a workaround, not a fix. They’re essentially saying, “Stop being weird about goblins,” instead of addressing why the models are goblin-obsessed in the first place. That’s fine for a production rollout, but it doesn’t inspire confidence in their ability to control model behavior at scale.

OpenAI’s blog post frames this as a transparency exercise — they’re explaining a known quirk rather than hiding it. That’s a step up from their usual opacity, but let’s be real: this only became public because Wired found the instruction. If journalists hadn’t poked around, would OpenAI have ever mentioned it? Probably not.

Still, I appreciate the candor in the post. They admit the problem existed, acknowledge it got worse, and describe their mitigation strategy. That’s more than most companies would do. The real question is whether this kind of emergent behavior is an isolated oddity or a sign of deeper issues in how these models generalize from training data. My money’s on the latter.

For now, if you’re using GPT-5.1 or later with the Nerdy personality, don’t be surprised if the model starts rambling about goblins. At least now you know why. And if you’re building on top of OpenAI’s APIs, maybe add your own “no fantasy creatures” guardrails — just in case.