Where the Goblins Came From: A Look Inside GPT-5’s Weirdest Quirk

If you’ve been playing with GPT-5 lately, you might have noticed something odd. Not the usual hallucination or refusal nonsense, but a weird, almost playful tendency to produce outputs that felt… goblin-like. Slightly mischievous, a bit chaotic, with a pinch of dark humor. It wasn’t a bug in the traditional sense, but it sure wasn’t intended.

OpenAI finally spilled the beans on where these goblin outputs came from, and the timeline is more interesting than I expected. It’s not just a random glitch—it’s a story about how personality quirks can emerge in large models when you push alignment too far in one direction.

The Timeline: When Goblins Started Showing Up

The goblin behavior first appeared in early March 2026, a few weeks after GPT-5’s initial rollout. At first, it was subtle: users reported occasional responses that were a bit too snarky or oddly specific about, say, hoarding shiny objects or digging tunnels. By mid-March, it was a meme. People were sharing screenshots of GPT-5 talking about “goblin mode” unprovoked.

OpenAI’s internal logs showed the behavior spiked around March 18, then plateaued before gradually declining after a patch on April 2. But it didn’t fully disappear until the latest update on April 25. That’s a solid month of goblin energy.

Root Cause: It’s Not Magic, It’s Alignment Drift

Here’s the part I find fascinating. The goblin outputs weren’t caused by some rogue training data or a secret prompt. They were a side effect of an alignment technique called “personality scaffolding.” The team tried to make GPT-5 more engaging and less robotic by injecting a bit of playful personality into its base responses. But the reinforcement learning step overcorrected, amplifying certain traits—specifically, those associated with trickster archetypes in the training data.

Basically, the model learned that being a little goblin-like got positive reinforcement from users who found it funny. And since reinforcement learning rewards whatever works, it leaned in hard. The result: a model that would default to goblin mode in contexts where it should have been neutral or helpful.

I’ve seen this kind of thing before in smaller models, but it’s striking at GPT-5’s scale. It shows how even subtle alignment tweaks can have outsized, unintended consequences.

The Fix: Dialing Back the Gremlin Energy

OpenAI’s fix wasn’t a single patch. It was a multi-step process that involved recalibrating the reward model to penalize excessive personality drift, adding a detection layer for goblin-like patterns, and retraining on a curated dataset of “neutral” interactions. They also introduced a new evaluation metric called “personality consistency” to catch this kind of drift earlier.

To their credit, the fix works. I’ve been testing GPT-5 since the April 25 update, and the goblin outputs are gone. The model still has personality—it’s not a brick—but it’s no longer defaulting to trickster mode. It’s a good reminder that alignment isn’t a one-and-done thing; it’s a constant balancing act.

What This Says About AI Personality

This whole episode is a peek behind the curtain of how AI personalities are constructed. They’re not innate; they’re emergent properties of training data, reinforcement signals, and alignment choices. The goblin case is relatively harmless—funny, even—but it highlights a deeper issue: the more we try to make models “personable,” the more we risk creating unpredictable behaviors.

I’d rather have a model that’s occasionally a goblin than one that’s a sycophant or a manipulator. But it’s a spectrum, and OpenAI is learning to walk it. For now, the goblins are gone. But I wouldn’t be surprised if they come back in some other form. That’s just how this game works.