OpenAI Finally Explains Why Its Models Won’t Stop Talking About Goblins

OpenAI finally addressed the weirdest AI behavior I’ve seen in a while: its models developed a strange obsession with goblins, gremlins, raccoons, trolls, ogres, and pigeons. Not in a fun, fantasy way — more like an annoying tic that kept popping up in responses.

It started with GPT-5.1, specifically when users chose the “Nerdy” personality option. The models would drop metaphors about these creatures unprompted. And it got worse with each update. Wired first broke the story, revealing OpenAI had actually hardcoded instructions telling the coding model to “never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures.”

Now OpenAI has published an explanation on their site, calling it a “strange habit” that emerged from training data quirks. Basically, the models learned to associate certain problem-solving patterns with fantasy creature references, likely because the training data included a lot of forum posts or documentation where people used those metaphors. The Nerdy personality amplified it because that mode tries to be more creative and reference-heavy.

This is honestly hilarious and slightly concerning. You’d think if you’re building a multi-billion dollar AI system, you’d catch something like this before shipping. But nope — it took a journalist’s report to force an admission. The fix? They apparently added explicit instructions to suppress those references, but that’s a band-aid, not a real solution. The underlying training data issue is still there.

I’ve seen similar quirks in other models — like one that kept talking about “the matrix” or another that couldn’t stop referencing “the singularity.” But goblins? That’s a new one. It shows how unpredictable these systems are. You can’t just throw data at them and expect clean behavior. The training process creates weird correlations we don’t fully understand.

OpenAI’s blog post is worth reading if you’re into the technical weeds, but the takeaway is simple: AI models are still glorified pattern-matching machines, and sometimes the patterns they match are ridiculous. Until we figure out how to train them without these artifacts, expect more bizarre behavior — and more emergency patches to ban specific topics.