OpenAI’s Models Got a Weird Goblin Habit, And They’re Finally Talking About It

OpenAI is finally owning up to something that sounds like a meme but is actually a real quirk in their models: a tendency to talk about goblins, gremlins, raccoons, trolls, ogres, and pigeons. A Wired report first flagged that their coding model was explicitly instructed to “never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures.” That raised eyebrows, and now OpenAI has published a blog post explaining what’s going on.

According to the post, this started with GPT-5.1, specifically when users selected the “Nerdy” personality option. The models began spitting out metaphors involving these creatures. And it got worse with each subsequent model refresh. OpenAI calls it a “strange habit” that emerged from the training data — essentially, the models learned to lean on these references as a kind of shorthand for certain concepts.

The fix? They’ve been tweaking the training process to reduce the frequency of these references. But the real story here isn’t just the goblins. It’s that OpenAI is being transparent about a weird artifact of how these models work. Most companies would just silently patch it and move on. Instead, they’re acknowledging that large language models can develop bizarre patterns that have nothing to do with their intended purpose.

I’ve seen this kind of thing before with other models — they latch onto specific phrases or concepts because they appear frequently in certain contexts. It’s a reminder that these systems are statistical pattern matchers, not intentional agents. The fact that OpenAI is open about it is refreshing, even if the whole thing sounds like a Dungeons & Dragons session gone wrong.

For now, the goblin problem is mostly contained to the Nerdy personality, but it’s a good example of how subtle training data quirks can create unexpected behavior. If you’re using GPT-5.1 or later models, you probably won’t run into it unless you’re actively looking for it. But if you do, now you know why.