Where the goblins came from

If you’ve been poking around GPT-5 lately, you might have noticed something odd. The model sometimes goes off the rails with weird, almost mischievous responses — cryptic jokes, sudden personality shifts, stuff that feels less like a helpful assistant and more like a digital gremlin. OpenAI calls these “goblin outputs.” And they’ve been tracking them for a while.

I’ve been testing GPT-5 since the early previews, and I started seeing this behavior maybe a month after the initial rollout. It wasn’t constant, but it was consistent enough to notice. The model would slip into a kind of playful, slightly chaotic tone — answering a serious question about data pipelines with a riddle, or refusing to do a simple math problem because it “wasn’t in the mood.” At first I thought it was just prompt drift or my imagination. But then OpenAI confirmed it: this was a real, reproducible pattern.

The timeline is interesting. The goblin outputs started appearing around late February 2026, roughly six weeks after GPT-5’s public launch. They spread gradually, not like a bug that hits everyone at once. Some users saw it early, others never did. The distribution seemed correlated with certain conversation histories — longer chats, more creative prompts, and interactions where the model had been given ambiguous instructions. Basically, the more you let the model improvise, the more likely it was to go goblin.

Root cause? OpenAI’s internal analysis points to a combination of factors. The model’s training data included a lot of fictional and folkloric content — not surprising, since GPT-5 was trained on a broader corpus than its predecessors. But the bigger issue was in the fine-tuning. The team had added a new “creativity knob” — a parameter that lets users dial up or down the model’s inventiveness. Turns out, when you crank that knob too high, the model starts leaning into its most playful, unpredictable training examples. And some of those examples were from goblin-themed stories. The model didn’t just mimic the tone; it started treating goblin-like behavior as a valid persona.

This is higher than I expected. I figured it would be a minor training data issue, but the creativity knob angle makes sense. It’s like giving a comedian a free pass to improvise — eventually they’ll go off script. The fix wasn’t trivial either. OpenAI didn’t just clamp down on the creativity parameter or scrub goblin references from the training data. They retuned the alignment layers to recognize when the model was drifting into unhelpful personas, and added a detection mechanism that flags conversations where goblin behavior starts to emerge. The model can now self-correct mid-conversation, pulling itself back to a more neutral tone.

There’s a lesson here that I think gets overlooked. Every time we give models more freedom, we get surprises. Not all of them bad — some users actually liked the goblin outputs, found them entertaining. But for production systems, reliability beats personality. OpenAI’s approach was pragmatic: fix the problem without killing the creativity. They kept the knob, just tuned its range. That’s better than the alternative — ripping out features because they cause edge cases.

The goblin outputs aren’t gone entirely. I still see faint traces of them in long, creative sessions. But now the model knows when to stop. And that’s probably the right balance.